program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func bucket_1024<ios18>(tensor<int32, [1, 1024]> input_ids, tensor<int32, [1, 1024]> position_ids, tensor<fp32, [1, 1024]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 1024]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 1024]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 1024]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 1024]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 1024]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 1024]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 1024, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 1024]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 1024]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 1024]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 1024, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 1024, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 1024, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 1024, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 1024, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 1024, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1024, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1024, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 1024, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 1024, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 1024, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 1024, 1024]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495408192))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 1024, 1024]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 1024, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 1024, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 1024, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 1024, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 1024, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 1024, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 1024, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 1024, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 1024, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 1024, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 1024, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 1024, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 1024, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 1024, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 1024, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 1024, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 1024, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 1024, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 1024, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 1024, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 1024, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 1024, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 1024, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 1024, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 1024, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 1024, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 1024, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 1024, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 1024, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 1024, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 1024, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 1024, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 1024, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 1024, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 1024, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 1024, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 1024, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 1024, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 1024, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 1024, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 1024, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 1024, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 1024, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 1024, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 1024, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 1024, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 1024, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 1024, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 1024, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 1024, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 1024, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 1024, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 1024, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 1024, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 1024, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 1024, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 1024, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 1024, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 1024, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 1024, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 1024, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 1024, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 1024, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 1024, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 1024, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 1024, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 1024, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 1024, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 1024, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 1024, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 1024, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 1024, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 1024, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 1024, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 1024, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 1024, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 1024, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 1024, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 1024, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 1024, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 1024, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 1024, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 1024, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 1024, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 1024, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 1024, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1024, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 1024, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 1024, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 1024, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 1024, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 1024, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 1024, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 1024, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 1024, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 1024, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 1024, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 1024, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 1024, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 1024, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 1024, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 1024, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 1024, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 1024, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 1024, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 1024, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 1024, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 1024, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 1024, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 1024, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 1024, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 1024, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 1024, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 1024, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 1024, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 1024, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 1024, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 1024, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 1024, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 1024, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 1024, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 1024, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 1024, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 1024, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 1024, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 1024, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 1024, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 1024, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 1024, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 1024, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 1024, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 1024, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 1024, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 1024, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 1024, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 1024, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 1024, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 1024, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 1024, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 1024, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 1024, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 1024, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 1024, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 1024, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 1024, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 1024, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 1024, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 1024, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 1024, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1024, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 1024, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 1024, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 1024, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 1024, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 1024, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 1024, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 1024, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 1024, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 1024, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 1024, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 1024, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 1024, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 1024, -1, 64])];
            tensor<fp16, [1, 1024, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 1024, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 1024, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 1024, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 1024, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 1024, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 1024, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 1024, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 1024, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 1024, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 1024, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 1024, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1024, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 1024, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 1024, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 1024, 64])];
            tensor<fp16, [1, 14, 1024, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 1024]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 1024, 1024]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 1024, 1024]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 1024, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 1024, -1])];
            tensor<fp16, [1, 1024, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 1024, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 1024, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 1024, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1024, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1024, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 1024, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1024, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1024, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1024, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1024, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 1024, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1024, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 1024, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1024]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 1024, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 1024, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_128<ios18>(tensor<int32, [1, 128]> input_ids, tensor<int32, [1, 128]> position_ids, tensor<fp32, [1, 128]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 128]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 128]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 128]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 128]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 128]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 128, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 128]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 128]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 128]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 128, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 128, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 128, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 128, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 128, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 128, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 128, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 128, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 128, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 128, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 128, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 128, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 128, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 128, 128]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495063936))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 128, 128]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 128, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 128, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 128, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 128, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 128, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 128, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 128, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 128, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 128, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 128, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 128, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 128, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 128, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 128, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 128, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 128, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 128, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 128, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 128, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 128, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 128, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 128, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 128, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 128, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 128, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 128, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 128, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 128, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 128, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 128, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 128, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 128, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 128, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 128, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 128, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 128, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 128, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 128, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 128, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 128, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 128, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 128, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 128, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 128, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 128, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 128, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 128, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 128, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 128, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 128, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 128, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 128, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 128, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 128, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 128, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 128, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 128, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 128, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 128, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 128, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 128, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 128, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 128, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 128, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 128, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 128, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 128, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 128, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 128, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 128, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 128, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 128, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 128, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 128, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 128, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 128, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 128, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 128, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 128, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 128, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 128, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 128, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 128, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 128, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 128, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 128, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 128, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 128, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 128, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 128, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 128, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 128, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 128, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 128, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 128, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 128, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 128, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 128, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 128, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 128, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 128, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 128, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 128, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 128, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 128, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 128, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 128, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 128, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 128, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 128, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 128, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 128, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 128, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 128, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 128, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 128, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 128, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 128, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 128, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 128, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 128, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 128, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 128, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 128, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 128, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 128, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 128, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 128, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 128, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 128, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 128, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 128, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 128, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 128, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 128, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 128, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 128, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 128, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 128, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 128, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 128, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 128, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 128, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 128, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 128, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 128, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 128, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 128, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 128, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 128, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 128, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 128, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 128, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 128, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 128, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 128, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 128, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 128, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 128, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 128, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 128, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 128, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 128, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 128, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 128, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 128, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 128, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 128, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 128, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 128, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 128, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 128, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 128, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 128, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 128, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 128, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 128, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 128, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 128, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 128, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 128, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 128, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 128, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 128, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 128, -1, 64])];
            tensor<fp16, [1, 128, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 128, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 128, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 128, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 128, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 128, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 128, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 128, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 128, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 128, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 128, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 128, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 128, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 128, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 128, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 128, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 128, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 128, 64])];
            tensor<fp16, [1, 14, 128, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 128, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 128]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 128, 128]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 128, 128]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 128, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 128, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 128, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 128, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 128, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 128, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 128, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 128, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 128, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 128, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 128, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 128, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 128, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 128]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 128, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 128, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_128_b4<ios18>(tensor<int32, [4, 128]> input_ids, tensor<int32, [4, 128]> position_ids, tensor<fp32, [4, 128]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [4, 128]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [4, 128]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [4, 128]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [4, 128]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [4, 128]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [4, 128]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [4, 128, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [4, 1, 128]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [4, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500651200)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 1, 128]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [4, 32, 128]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 128, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 128, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [4, 128, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [4, 128, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [4, 128, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [4, 128, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [4, 128, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [4, 128, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 128, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 128, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 14, 128, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 2, 128, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [4, 14, 128, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [4, 1, 128, 128]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4, 1, 128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500672128))), scale = tensor<fp16, [4, 1, 1, 1]>([[[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [4, 14, 128, 128]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [4, 128, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [4, 128, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [4, 128, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [4, 128, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [4, 128, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [4, 128, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [4, 128, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [4, 128, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [4, 14, 128, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [4, 2, 128, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [4, 128, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [4, 128, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [4, 128, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [4, 128, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [4, 128, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [4, 128, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [4, 14, 128, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 2, 128, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 128, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [4, 128, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [4, 128, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [4, 128, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [4, 128, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [4, 128, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 14, 128, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [4, 2, 128, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [4, 128, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [4, 128, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [4, 128, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [4, 128, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [4, 128, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [4, 128, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [4, 14, 128, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [4, 2, 128, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 128, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [4, 128, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [4, 128, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [4, 128, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [4, 128, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [4, 128, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 14, 128, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 2, 128, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [4, 128, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [4, 128, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [4, 128, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [4, 128, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [4, 128, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [4, 128, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [4, 14, 128, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [4, 2, 128, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [4, 128, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [4, 128, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [4, 128, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [4, 128, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [4, 128, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [4, 128, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [4, 14, 128, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 2, 128, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 128, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [4, 128, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [4, 128, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [4, 128, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [4, 128, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [4, 128, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 14, 128, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [4, 2, 128, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [4, 128, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [4, 128, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [4, 128, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [4, 128, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [4, 128, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [4, 128, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [4, 14, 128, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [4, 2, 128, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 128, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [4, 128, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [4, 128, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [4, 128, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [4, 128, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [4, 128, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 14, 128, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 2, 128, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [4, 128, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [4, 128, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [4, 128, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [4, 128, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [4, 128, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [4, 128, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [4, 14, 128, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [4, 2, 128, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [4, 128, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [4, 128, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [4, 128, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [4, 128, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [4, 128, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [4, 128, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [4, 14, 128, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 2, 128, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 128, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [4, 128, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [4, 128, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [4, 128, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [4, 128, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [4, 128, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 14, 128, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [4, 2, 128, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [4, 128, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [4, 128, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [4, 128, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [4, 128, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [4, 128, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [4, 128, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [4, 14, 128, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [4, 2, 128, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 128, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [4, 128, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [4, 128, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [4, 128, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [4, 128, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [4, 128, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 14, 128, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 2, 128, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [4, 128, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [4, 128, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [4, 128, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [4, 128, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [4, 128, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [4, 128, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [4, 14, 128, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [4, 2, 128, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [4, 128, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [4, 128, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [4, 128, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [4, 128, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [4, 128, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [4, 128, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [4, 14, 128, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 2, 128, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 128, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [4, 128, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [4, 128, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [4, 128, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [4, 128, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [4, 128, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 14, 128, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [4, 2, 128, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [4, 128, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [4, 128, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [4, 128, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [4, 128, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [4, 128, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [4, 128, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [4, 14, 128, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [4, 2, 128, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 128, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [4, 128, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [4, 128, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [4, 128, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [4, 128, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [4, 128, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 14, 128, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 2, 128, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [4, 128, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [4, 128, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [4, 128, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [4, 128, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [4, 128, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [4, 128, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [4, 14, 128, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [4, 2, 128, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [4, 128, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [4, 128, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [4, 128, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [4, 128, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [4, 128, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [4, 128, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [4, 14, 128, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 2, 128, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 128, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [4, 128, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [4, 128, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [4, 128, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [4, 128, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [4, 128, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([4, 128, -1, 64])];
            tensor<fp16, [4, 128, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 128, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 14, 128, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([4, 14, 128, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 128, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 128, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 128, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [4, 2, 128, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([4, 2, 128, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 128, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([4, 2, 128, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 128, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 128, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 128, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 128, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [4, 2, 1, 128, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 128, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([4, 14, 128, 64])];
            tensor<fp16, [4, 14, 128, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [4, 14, 128, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 128]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [4, 14, 128, 128]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 128, 128]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 128, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([4, 128, -1])];
            tensor<fp16, [4, 128, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [4, 128, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [4, 128, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [4, 128, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 128, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [4, 128, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 128, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 128, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [4, 128, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 128, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 128, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 128, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 128, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [4, 128, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 128, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [4, 128, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [4, 128, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 128]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [4, 128, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [4, 128, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [4, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [4, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [4, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [4, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [4, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [4, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_2048<ios18>(tensor<int32, [1, 2048]> input_ids, tensor<int32, [1, 2048]> position_ids, tensor<fp32, [1, 2048]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 2048]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 2048]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 2048]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 2048]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 2048]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 2048]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 2048, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 2048]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 2048]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 2048]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2048, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2048, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 2048, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 2048, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 2048, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 2048, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 2048, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2048, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 2048, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 2048, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 2048, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 2048, 2048]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 2048, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(496456832))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 2048, 2048]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 2048, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 2048, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 2048, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 2048, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 2048, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 2048, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 2048, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 2048, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 2048, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 2048, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 2048, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 2048, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 2048, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 2048, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 2048, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 2048, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 2048, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 2048, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 2048, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 2048, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 2048, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 2048, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 2048, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 2048, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 2048, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 2048, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 2048, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 2048, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 2048, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 2048, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 2048, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 2048, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 2048, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 2048, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 2048, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 2048, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 2048, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 2048, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 2048, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 2048, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 2048, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 2048, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 2048, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 2048, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 2048, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 2048, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 2048, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 2048, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 2048, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 2048, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 2048, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 2048, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 2048, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 2048, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 2048, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 2048, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 2048, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 2048, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 2048, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 2048, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 2048, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 2048, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 2048, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 2048, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 2048, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 2048, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 2048, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 2048, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 2048, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 2048, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 2048, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 2048, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 2048, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 2048, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 2048, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 2048, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 2048, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 2048, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 2048, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 2048, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 2048, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 2048, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 2048, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 2048, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 2048, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 2048, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 2048, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 2048, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 2048, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 2048, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 2048, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 2048, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 2048, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 2048, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 2048, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 2048, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 2048, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 2048, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 2048, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 2048, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 2048, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 2048, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 2048, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 2048, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 2048, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 2048, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 2048, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 2048, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 2048, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 2048, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 2048, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 2048, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 2048, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 2048, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 2048, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 2048, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 2048, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 2048, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 2048, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 2048, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 2048, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 2048, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 2048, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 2048, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 2048, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 2048, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 2048, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 2048, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 2048, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 2048, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 2048, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 2048, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 2048, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 2048, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 2048, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 2048, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 2048, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 2048, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 2048, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 2048, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 2048, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 2048, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 2048, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 2048, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 2048, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 2048, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 2048, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 2048, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 2048, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 2048, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 2048, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 2048, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 2048, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 2048, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 2048, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 2048, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 2048, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 2048, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 2048, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 2048, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 2048, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 2048, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 2048, -1, 64])];
            tensor<fp16, [1, 2048, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 2048, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 2048, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 2048, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 2048, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 2048, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 2048, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 2048, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 2048, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 2048, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 2048, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 2048, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 2048, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 2048, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 2048, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 2048, 64])];
            tensor<fp16, [1, 14, 2048, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 2048]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 2048, 2048]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 2048, 2048]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 2048, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 2048, -1])];
            tensor<fp16, [1, 2048, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 2048, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 2048, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 2048, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2048, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2048, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 2048, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2048, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2048, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2048, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2048, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 2048, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2048, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 2048, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 2048]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 2048, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 2048, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_256<ios18>(tensor<int32, [1, 256]> input_ids, tensor<int32, [1, 256]> position_ids, tensor<fp32, [1, 256]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 256]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 256]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 256]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 256]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 256]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 256]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 256, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 256]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 256]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 256]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 256, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 256, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 256, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 256, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 256, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 256, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 256, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 256, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 256, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 256, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 256, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 256, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 256, 256]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495080384))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 256, 256]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 256, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 256, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 256, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 256, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 256, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 256, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 256, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 256, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 256, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 256, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 256, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 256, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 256, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 256, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 256, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 256, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 256, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 256, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 256, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 256, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 256, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 256, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 256, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 256, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 256, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 256, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 256, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 256, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 256, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 256, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 256, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 256, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 256, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 256, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 256, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 256, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 256, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 256, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 256, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 256, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 256, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 256, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 256, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 256, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 256, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 256, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 256, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 256, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 256, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 256, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 256, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 256, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 256, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 256, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 256, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 256, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 256, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 256, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 256, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 256, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 256, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 256, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 256, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 256, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 256, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 256, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 256, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 256, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 256, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 256, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 256, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 256, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 256, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 256, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 256, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 256, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 256, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 256, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 256, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 256, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 256, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 256, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 256, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 256, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 256, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 256, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 256, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 256, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 256, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 256, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 256, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 256, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 256, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 256, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 256, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 256, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 256, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 256, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 256, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 256, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 256, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 256, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 256, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 256, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 256, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 256, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 256, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 256, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 256, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 256, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 256, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 256, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 256, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 256, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 256, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 256, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 256, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 256, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 256, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 256, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 256, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 256, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 256, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 256, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 256, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 256, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 256, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 256, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 256, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 256, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 256, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 256, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 256, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 256, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 256, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 256, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 256, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 256, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 256, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 256, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 256, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 256, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 256, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 256, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 256, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 256, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 256, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 256, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 256, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 256, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 256, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 256, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 256, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 256, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 256, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 256, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 256, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 256, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 256, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 256, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 256, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 256, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 256, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 256, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 256, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 256, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 256, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 256, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 256, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 256, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 256, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 256, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 256, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 256, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 256, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 256, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 256, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 256, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 256, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 256, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 256, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 256, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 256, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 256, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 256, -1, 64])];
            tensor<fp16, [1, 256, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 256, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 256, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 256, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 256, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 256, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 256, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 256, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 256, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 256, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 256, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 256, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 256, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 256, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 256, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 256, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 256, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 256, 64])];
            tensor<fp16, [1, 14, 256, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 256, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 256]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 256, 256]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 256, 256]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 256, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 256, -1])];
            tensor<fp16, [1, 256, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 256, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 256, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 256, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 256, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 256, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 256, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 256, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 256, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 256, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 256, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 256, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 256, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 256, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 256, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 256, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 256, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 256, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 256, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_32<ios18>(tensor<int32, [1, 32]> input_ids, tensor<int32, [1, 32]> position_ids, tensor<fp32, [1, 32]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 32]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 32]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 32]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 32]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 32]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 32]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 32, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 32]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 32]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 32]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 32, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 32, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 32, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 32, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 32, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 32, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 32, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 32, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 32, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 32, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 32, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 32, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 32, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 32, 32]> mul_0_to_fp16 = const()[name = string("mul_0_to_fp16"), val = tensor<fp16, [1, 1, 32, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137478080)))];
            tensor<fp16, [1, 14, 32, 32]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 32, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 32, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 32, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 32, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 32, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 32, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 32, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 32, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 32, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 32, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 32, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 32, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 32, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 32, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 32, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 32, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 32, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 32, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 32, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 32, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 32, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 32, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 32, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 32, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 32, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 32, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 32, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 32, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 32, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 32, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 32, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 32, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 32, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 32, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 32, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 32, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 32, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 32, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 32, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 32, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 32, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 32, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 32, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 32, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 32, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 32, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 32, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 32, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 32, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 32, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 32, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 32, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 32, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 32, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 32, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 32, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 32, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 32, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 32, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 32, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 32, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 32, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 32, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 32, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 32, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 32, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 32, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 32, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 32, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 32, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 32, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 32, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 32, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 32, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 32, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 32, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 32, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 32, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 32, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 32, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 32, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 32, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 32, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 32, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 32, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 32, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 32, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 32, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 32, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 32, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 32, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 32, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 32, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 32, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 32, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 32, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 32, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 32, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 32, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 32, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 32, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 32, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 32, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 32, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 32, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 32, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 32, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 32, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 32, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 32, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 32, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 32, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 32, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 32, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 32, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 32, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 32, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 32, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 32, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 32, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 32, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 32, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 32, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 32, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 32, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 32, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 32, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 32, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 32, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 32, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 32, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 32, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 32, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 32, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 32, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 32, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 32, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 32, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 32, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 32, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 32, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 32, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 32, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 32, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 32, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 32, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 32, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 32, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 32, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 32, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 32, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 32, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 32, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 32, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 32, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 32, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 32, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 32, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 32, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 32, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 32, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 32, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 32, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 32, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 32, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 32, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 32, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 32, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 32, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 32, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 32, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 32, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 32, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 32, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 32, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 32, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 32, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 32, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 32, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 32, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 32, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 32, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 32, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 32, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 32, -1, 64])];
            tensor<fp16, [1, 32, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 32, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 32, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 32, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 32, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 32, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 32, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 32, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 32, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 32, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 32, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 32, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 32, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 32, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 32, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 32, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 32, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 32, 64])];
            tensor<fp16, [1, 14, 32, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 32, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 32]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 32, 32]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 32, 32]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 32, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 32, -1])];
            tensor<fp16, [1, 32, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 32, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 32, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 32, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 32, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 32, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 32, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 32, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 32, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 32, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 32, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 32, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 32, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 32, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 32, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 32, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 32, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 32]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 32, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 32, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_32_b4<ios18>(tensor<int32, [4, 32]> input_ids, tensor<int32, [4, 32]> position_ids, tensor<fp32, [4, 32]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [4, 32]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [4, 32]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [4, 32]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [4, 32]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [4, 32]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [4, 32]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [4, 32, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [4, 1, 32]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [4, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500651200)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 1, 32]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [4, 32, 32]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 32, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 32, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [4, 32, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [4, 32, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [4, 32, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [4, 32, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [4, 32, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [4, 32, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 32, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 32, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 14, 32, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 2, 32, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [4, 14, 32, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [4, 1, 32, 32]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4, 1, 32, 32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500651520))), scale = tensor<fp16, [4, 1, 1, 1]>([[[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [4, 14, 32, 32]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [4, 32, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [4, 32, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [4, 32, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [4, 32, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [4, 32, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [4, 32, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [4, 32, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [4, 32, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [4, 14, 32, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [4, 2, 32, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [4, 32, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [4, 32, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [4, 32, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [4, 32, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [4, 32, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [4, 32, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [4, 14, 32, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 2, 32, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 32, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [4, 32, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [4, 32, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [4, 32, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [4, 32, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [4, 32, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 14, 32, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [4, 2, 32, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [4, 32, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [4, 32, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [4, 32, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [4, 32, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [4, 32, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [4, 32, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [4, 14, 32, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [4, 2, 32, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 32, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [4, 32, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [4, 32, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [4, 32, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [4, 32, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [4, 32, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 14, 32, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 2, 32, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [4, 32, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [4, 32, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [4, 32, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [4, 32, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [4, 32, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [4, 32, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [4, 14, 32, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [4, 2, 32, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [4, 32, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [4, 32, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [4, 32, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [4, 32, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [4, 32, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [4, 32, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [4, 14, 32, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 2, 32, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 32, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [4, 32, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [4, 32, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [4, 32, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [4, 32, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [4, 32, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 14, 32, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [4, 2, 32, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [4, 32, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [4, 32, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [4, 32, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [4, 32, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [4, 32, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [4, 32, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [4, 14, 32, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [4, 2, 32, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 32, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [4, 32, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [4, 32, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [4, 32, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [4, 32, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [4, 32, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 14, 32, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 2, 32, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [4, 32, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [4, 32, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [4, 32, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [4, 32, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [4, 32, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [4, 32, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [4, 14, 32, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [4, 2, 32, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [4, 32, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [4, 32, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [4, 32, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [4, 32, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [4, 32, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [4, 32, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [4, 14, 32, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 2, 32, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 32, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [4, 32, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [4, 32, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [4, 32, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [4, 32, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [4, 32, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 14, 32, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [4, 2, 32, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [4, 32, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [4, 32, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [4, 32, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [4, 32, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [4, 32, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [4, 32, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [4, 14, 32, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [4, 2, 32, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 32, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [4, 32, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [4, 32, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [4, 32, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [4, 32, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [4, 32, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 14, 32, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 2, 32, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [4, 32, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [4, 32, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [4, 32, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [4, 32, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [4, 32, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [4, 32, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [4, 14, 32, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [4, 2, 32, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [4, 32, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [4, 32, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [4, 32, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [4, 32, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [4, 32, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [4, 32, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [4, 14, 32, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 2, 32, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 32, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [4, 32, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [4, 32, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [4, 32, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [4, 32, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [4, 32, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 14, 32, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [4, 2, 32, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [4, 32, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [4, 32, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [4, 32, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [4, 32, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [4, 32, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [4, 32, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [4, 14, 32, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [4, 2, 32, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 32, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [4, 32, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [4, 32, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [4, 32, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [4, 32, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [4, 32, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 14, 32, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 2, 32, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [4, 32, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [4, 32, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [4, 32, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [4, 32, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [4, 32, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [4, 32, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [4, 14, 32, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [4, 2, 32, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [4, 32, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [4, 32, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [4, 32, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [4, 32, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [4, 32, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [4, 32, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [4, 14, 32, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 2, 32, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 32, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [4, 32, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [4, 32, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [4, 32, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [4, 32, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [4, 32, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([4, 32, -1, 64])];
            tensor<fp16, [4, 32, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 32, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 14, 32, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([4, 14, 32, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 32, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 32, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 32, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [4, 2, 32, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([4, 2, 32, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 32, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([4, 2, 32, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 32, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 32, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 32, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 32, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [4, 2, 1, 32, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 32, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([4, 14, 32, 64])];
            tensor<fp16, [4, 14, 32, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [4, 14, 32, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 32]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [4, 14, 32, 32]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 32, 32]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 32, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [4, 32, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [4, 32, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [4, 32, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 32, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [4, 32, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 32, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 32, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [4, 32, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 32, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 32, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 32, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 32, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [4, 32, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 32, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [4, 32, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [4, 32, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 32]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [4, 32, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [4, 32, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [4, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [4, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [4, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [4, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [4, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [4, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_512<ios18>(tensor<int32, [1, 512]> input_ids, tensor<int32, [1, 512]> position_ids, tensor<fp32, [1, 512]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 512]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 512]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 512]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 512]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 512]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 512]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 512]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 512]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 512, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 512, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 512, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 512, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 512, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 512, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 512, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 512, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 512, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 512, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 512, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 512, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 512, 512]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495145984))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 512, 512]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 512, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 512, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 512, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 512, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 512, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 512, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 512, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 512, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 512, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 512, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 512, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 512, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 512, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 512, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 512, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 512, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 512, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 512, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 512, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 512, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 512, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 512, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 512, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 512, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 512, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 512, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 512, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 512, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 512, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 512, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 512, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 512, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 512, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 512, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 512, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 512, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 512, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 512, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 512, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 512, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 512, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 512, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 512, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 512, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 512, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 512, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 512, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 512, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 512, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 512, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 512, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 512, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 512, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 512, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 512, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 512, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 512, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 512, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 512, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 512, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 512, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 512, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 512, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 512, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 512, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 512, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 512, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 512, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 512, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 512, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 512, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 512, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 512, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 512, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 512, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 512, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 512, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 512, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 512, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 512, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 512, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 512, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 512, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 512, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 512, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 512, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 512, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 512, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 512, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 512, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 512, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 512, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 512, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 512, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 512, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 512, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 512, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 512, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 512, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 512, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 512, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 512, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 512, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 512, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 512, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 512, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 512, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 512, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 512, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 512, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 512, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 512, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 512, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 512, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 512, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 512, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 512, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 512, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 512, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 512, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 512, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 512, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 512, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 512, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 512, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 512, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 512, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 512, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 512, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 512, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 512, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 512, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 512, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 512, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 512, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 512, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 512, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 512, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 512, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 512, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 512, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 512, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 512, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 512, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 512, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 512, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 512, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 512, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 512, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 512, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 512, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 512, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 512, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 512, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 512, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 512, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 512, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 512, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 512, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 512, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 512, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 512, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 512, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 512, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 512, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 512, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 512, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 512, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 512, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 512, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 512, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 512, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 512, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 512, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 512, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 512, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 512, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 512, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 512, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 512, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 512, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 512, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 512, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 512, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 512, -1, 64])];
            tensor<fp16, [1, 512, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 512, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 512, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 512, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 512, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 512, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 512, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 512, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 512, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 512, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 512, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 512, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 512, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 512, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 512, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 512, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 512, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 512, 64])];
            tensor<fp16, [1, 14, 512, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 512, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 512]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 512, 512]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 512, 512]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 512, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 512, -1])];
            tensor<fp16, [1, 512, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 512, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 512, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 512, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 512, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 512, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 512, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 512, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 512, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 512, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 512, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 512, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 512, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 512]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 512, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 512, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_64<ios18>(tensor<int32, [1, 64]> input_ids, tensor<int32, [1, 64]> position_ids, tensor<fp32, [1, 64]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 64]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 64]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [1, 64]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 64]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 64]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [1, 64]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 64, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 64]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438720)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 64]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [1, 32, 64]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 64, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 64, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [1, 64, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [1, 64, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [1, 64, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [1, 64, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [1, 64, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [1, 64, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 64, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 64, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 14, 64, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 2, 64, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 14, 64, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [1, 1, 64, 64]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 1, 64, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495059776))), scale = tensor<fp16, [1, 1, 1, 1]>([[[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [1, 14, 64, 64]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 64, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [1, 64, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [1, 64, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [1, 64, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [1, 64, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [1, 64, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [1, 64, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [1, 64, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 14, 64, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 2, 64, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 64, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [1, 64, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [1, 64, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [1, 64, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [1, 64, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [1, 64, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 14, 64, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 2, 64, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 64, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [1, 64, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [1, 64, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [1, 64, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [1, 64, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [1, 64, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 14, 64, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 2, 64, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 64, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [1, 64, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [1, 64, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [1, 64, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [1, 64, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [1, 64, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 14, 64, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 2, 64, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 64, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [1, 64, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [1, 64, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [1, 64, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [1, 64, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [1, 64, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 14, 64, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 2, 64, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 64, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [1, 64, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [1, 64, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [1, 64, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [1, 64, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [1, 64, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 14, 64, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 2, 64, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 64, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [1, 64, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [1, 64, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [1, 64, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [1, 64, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [1, 64, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 14, 64, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 2, 64, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 64, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [1, 64, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [1, 64, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [1, 64, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [1, 64, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [1, 64, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 14, 64, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 2, 64, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 64, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [1, 64, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [1, 64, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [1, 64, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [1, 64, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [1, 64, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 14, 64, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 2, 64, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 64, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [1, 64, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [1, 64, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [1, 64, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [1, 64, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [1, 64, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 14, 64, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 2, 64, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 64, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [1, 64, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [1, 64, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [1, 64, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [1, 64, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [1, 64, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 14, 64, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 2, 64, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 64, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [1, 64, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [1, 64, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [1, 64, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [1, 64, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [1, 64, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 14, 64, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 2, 64, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 64, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [1, 64, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [1, 64, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [1, 64, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [1, 64, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [1, 64, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 14, 64, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 2, 64, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 64, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [1, 64, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [1, 64, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [1, 64, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [1, 64, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [1, 64, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 14, 64, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 2, 64, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 64, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [1, 64, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [1, 64, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [1, 64, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [1, 64, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [1, 64, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 14, 64, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 2, 64, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 64, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [1, 64, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [1, 64, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [1, 64, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [1, 64, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [1, 64, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 14, 64, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 2, 64, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 64, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [1, 64, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [1, 64, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [1, 64, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [1, 64, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [1, 64, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 14, 64, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 2, 64, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 64, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [1, 64, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [1, 64, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [1, 64, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [1, 64, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [1, 64, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 14, 64, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 2, 64, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 64, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [1, 64, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [1, 64, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [1, 64, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [1, 64, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [1, 64, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 14, 64, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 2, 64, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 64, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [1, 64, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [1, 64, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [1, 64, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [1, 64, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [1, 64, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 14, 64, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 2, 64, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 64, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [1, 64, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [1, 64, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [1, 64, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [1, 64, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [1, 64, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 14, 64, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 2, 64, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 64, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [1, 64, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [1, 64, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [1, 64, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [1, 64, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [1, 64, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 14, 64, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 2, 64, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 64, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [1, 64, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [1, 64, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [1, 64, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [1, 64, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [1, 64, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([1, 64, -1, 64])];
            tensor<fp16, [1, 64, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 14, 64, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 14, 64, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 14, 64, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 14, 64, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 14, 64, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 14, 64, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 2, 64, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 2, 64, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2, 64, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 2, 64, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2, 64, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2, 64, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2, 64, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2, 64, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 2, 1, 64, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [1, 2, 7, 64, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([1, 14, 64, 64])];
            tensor<fp16, [1, 14, 64, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [1, 14, 64, 64]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 14, 64, 64]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 14, 64, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([1, 64, -1])];
            tensor<fp16, [1, 64, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 64, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [1, 64, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 64, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 64, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 64, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 64, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [1, 64, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [1, 64, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 64, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 64, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 64, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 64, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [1, 64, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 64, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [1, 64, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [1, 64, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 64]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [1, 64, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [1, 64, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [1, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [1, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [1, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
    func bucket_64_b4<ios18>(tensor<int32, [4, 64]> input_ids, tensor<int32, [4, 64]> position_ids, tensor<fp32, [4, 64]> selector) {
            int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
            bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 896]> lm_embed_tokens_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [151936, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [151936, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136134784))))[name = string("lm_embed_tokens_weight_to_fp16_quantized")];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [4, 64]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [4, 64]> add_24 = add(x = input_ids, y = slice_by_index_0)[name = string("add_24")];
            tensor<int32, [4, 64]> select_0 = select(a = input_ids, b = add_24, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [4, 64]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [4, 64]> add_0 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0")];
            tensor<int32, [4, 64]> select_0_1 = select(a = select_0, b = add_0, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [4, 64, 896]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0_1, validate_indices = inputs_embeds_validate_indices_0, x = lm_embed_tokens_weight_to_fp16_quantized)[name = string("inputs_embeds_cast_fp16")];
            bool var_12 = const()[name = string("op_12"), val = bool(true)];
            int32 var_20 = const()[name = string("op_20"), val = int32(-1)];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [4, 1, 64]> var_108 = expand_dims(axes = var_108_axes_0, x = position_ids)[name = string("op_108")];
            bool var_113_transpose_x_0 = const()[name = string("op_113_transpose_x_0"), val = bool(false)];
            bool var_113_transpose_y_0 = const()[name = string("op_113_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 32, 1]> const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = tensor<fp16, [4, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500651200)))];
            string position_ids_expanded_1_to_fp16_dtype_0 = const()[name = string("position_ids_expanded_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 1, 64]> var_108_to_fp16 = cast(dtype = position_ids_expanded_1_to_fp16_dtype_0, x = var_108)[name = string("cast_2")];
            tensor<fp16, [4, 32, 64]> var_113_cast_fp16 = matmul(transpose_x = var_113_transpose_x_0, transpose_y = var_113_transpose_y_0, x = const_9_to_fp16, y = var_108_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            bool emb_interleave_0 = const()[name = string("emb_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 64, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_113_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [4, 64, 64]> emb_cast_fp16 = concat(axis = var_20, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_cast_fp16")];
            tensor<fp16, [4, 64, 64]> var_117_cast_fp16 = cos(x = emb_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<fp16, [4, 64, 64]> var_120_cast_fp16 = sin(x = emb_cast_fp16)[name = string("op_120_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_133_cast_fp16 = abs(x = inputs_embeds_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("op_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_135_cast_fp16 = reduce_max(axes = var_135_axes_0, keep_dims = var_12, x = var_133_cast_fp16)[name = string("op_135_cast_fp16")];
            fp16 var_11_to_fp16 = const()[name = string("op_11_to_fp16"), val = fp16(0x1.1p-20)];
            fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_0_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_10_to_fp16, x = var_135_cast_fp16)[name = string("clip_0_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_1_cast_fp16 = real_div(x = inputs_embeds_cast_fp16, y = clip_0_cast_fp16)[name = string("xn_1_cast_fp16")];
            fp16 var_23_promoted_to_fp16 = const()[name = string("op_23_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_138_cast_fp16 = pow(x = xn_1_cast_fp16, y = var_23_promoted_to_fp16)[name = string("op_138_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_138_cast_fp16)[name = string("var_1_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_141_cast_fp16 = mul(x = clip_0_cast_fp16, y = clip_0_cast_fp16)[name = string("op_141_cast_fp16")];
            fp32 var_142_epsilon_0 = const()[name = string("op_142_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_142_cast_fp16 = inverse(epsilon = var_142_epsilon_0, x = var_141_cast_fp16)[name = string("op_142_cast_fp16")];
            fp16 var_143_to_fp16 = const()[name = string("op_143_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_144_cast_fp16 = mul(x = var_142_cast_fp16, y = var_143_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_145_cast_fp16 = add(x = var_1_cast_fp16, y = var_144_cast_fp16)[name = string("op_145_cast_fp16")];
            fp32 var_146_epsilon_0 = const()[name = string("op_146_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_146_cast_fp16 = rsqrt(epsilon = var_146_epsilon_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_1_cast_fp16 = mul(x = xn_1_cast_fp16, y = var_146_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136438848)))];
            tensor<fp16, [4, 64, 896]> hidden_states_1_cast_fp16 = mul(x = lm_layers_0_input_layernorm_weight_to_fp16, y = out_1_cast_fp16)[name = string("hidden_states_1_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136440704))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137243584))))[name = string("lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137245440)))];
            tensor<fp16, [4, 64, 896]> linear_0_cast_fp16 = linear(bias = lm_layers_0_self_attn_q_proj_bias_to_fp16, weight = lm_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_158 = const()[name = string("op_158"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_159_cast_fp16 = reshape(shape = var_158, x = linear_0_cast_fp16)[name = string("op_159_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = string("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137247296))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362048))))[name = string("lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362368)))];
            tensor<fp16, [4, 64, 128]> linear_1_cast_fp16 = linear(bias = lm_layers_0_self_attn_k_proj_bias_to_fp16, weight = lm_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_164 = const()[name = string("op_164"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_165_cast_fp16 = reshape(shape = var_164, x = linear_1_cast_fp16)[name = string("op_165_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = string("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137362688))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477440))))[name = string("lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137477760)))];
            tensor<fp16, [4, 64, 128]> linear_2_cast_fp16 = linear(bias = lm_layers_0_self_attn_v_proj_bias_to_fp16, weight = lm_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_171_cast_fp16 = reshape(shape = var_170, x = linear_2_cast_fp16)[name = string("op_171_cast_fp16")];
            tensor<int32, [4]> hidden_states_7_perm_0 = const()[name = string("hidden_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_5_axes_0 = const()[name = string("cos_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 64, 64]> cos_5_cast_fp16 = expand_dims(axes = cos_5_axes_0, x = var_117_cast_fp16)[name = string("cos_5_cast_fp16")];
            tensor<int32, [1]> sin_5_axes_0 = const()[name = string("sin_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 64, 64]> sin_5_cast_fp16 = expand_dims(axes = sin_5_axes_0, x = var_120_cast_fp16)[name = string("sin_5_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_159_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [4, 14, 64, 64]> var_175_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_175_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_186_cast_fp16")];
            bool var_188_interleave_0 = const()[name = string("op_188_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_188_cast_fp16 = concat(axis = var_20, interleave = var_188_interleave_0, values = (var_186_cast_fp16, x1_1_cast_fp16))[name = string("op_188_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_189_cast_fp16 = mul(x = var_188_cast_fp16, y = sin_5_cast_fp16)[name = string("op_189_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_1_cast_fp16 = add(x = var_175_cast_fp16, y = var_189_cast_fp16)[name = string("query_1_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_165_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [4, 2, 64, 64]> var_191_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_5_cast_fp16)[name = string("op_191_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_202_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_202_cast_fp16")];
            bool var_204_interleave_0 = const()[name = string("op_204_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_204_cast_fp16 = concat(axis = var_20, interleave = var_204_interleave_0, values = (var_202_cast_fp16, x1_3_cast_fp16))[name = string("op_204_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = sin_5_cast_fp16)[name = string("op_205_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_3_cast_fp16 = add(x = var_191_cast_fp16, y = var_205_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<int32, [1]> var_214_axes_0 = const()[name = string("op_214_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_214_cast_fp16 = expand_dims(axes = var_214_axes_0, x = hidden_states_3_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<int32, [5]> hidden_states_5_reps_0 = const()[name = string("hidden_states_5_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_5_cast_fp16 = tile(reps = hidden_states_5_reps_0, x = var_214_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
            tensor<int32, [4]> var_222 = const()[name = string("op_222"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_1_cast_fp16 = reshape(shape = var_222, x = hidden_states_5_cast_fp16)[name = string("key_1_cast_fp16")];
            tensor<int32, [1]> var_231_axes_0 = const()[name = string("op_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_7_cast_fp16 = transpose(perm = hidden_states_7_perm_0, x = var_171_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_231_cast_fp16 = expand_dims(axes = var_231_axes_0, x = hidden_states_7_cast_fp16)[name = string("op_231_cast_fp16")];
            tensor<int32, [5]> hidden_states_9_reps_0 = const()[name = string("hidden_states_9_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_9_cast_fp16 = tile(reps = hidden_states_9_reps_0, x = var_231_cast_fp16)[name = string("hidden_states_9_cast_fp16")];
            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_1_cast_fp16 = reshape(shape = var_239, x = hidden_states_9_cast_fp16)[name = string("value_1_cast_fp16")];
            fp16 var_13_to_fp16 = const()[name = string("op_13_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [4, 14, 64, 64]> mul_1_cast_fp16 = mul(x = query_1_cast_fp16, y = var_13_to_fp16)[name = string("mul_1_cast_fp16")];
            bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)];
            bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = key_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<fp16, [4, 1, 64, 64]> mul_0_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4, 1, 64, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500655680))), scale = tensor<fp16, [4, 1, 1, 1]>([[[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]], [[[0x1.d88p+7]]]]))[name = string("mul_0_to_fp16_quantized")];
            tensor<fp16, [4, 14, 64, 64]> add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_0_cast_fp16")];
            int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = add_0_cast_fp16)[name = string("softmax_0_cast_fp16")];
            bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)];
            bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = value_1_cast_fp16)[name = string("attn_output_1_cast_fp16")];
            tensor<int32, [4]> var_242_perm_0 = const()[name = string("op_242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_242_cast_fp16 = transpose(perm = var_242_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [4, 64, 896]> var_245_cast_fp16 = reshape(shape = var_244, x = var_242_cast_fp16)[name = string("op_245_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137480192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138283072))))[name = string("lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> linear_3_bias_0_to_fp16 = const()[name = string("linear_3_bias_0_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138284928)))];
            tensor<fp16, [4, 64, 896]> linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = var_245_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_1_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_251_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_251_cast_fp16")];
            tensor<int32, [1]> var_253_axes_0 = const()[name = string("op_253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_253_cast_fp16 = reduce_max(axes = var_253_axes_0, keep_dims = var_12, x = var_251_cast_fp16)[name = string("op_253_cast_fp16")];
            fp16 const_27_to_fp16 = const()[name = string("const_27_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_1_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_27_to_fp16, x = var_253_cast_fp16)[name = string("clip_1_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = clip_1_cast_fp16)[name = string("xn_3_cast_fp16")];
            fp16 var_23_promoted_1_to_fp16 = const()[name = string("op_23_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_256_cast_fp16 = pow(x = xn_3_cast_fp16, y = var_23_promoted_1_to_fp16)[name = string("op_256_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_256_cast_fp16)[name = string("var_3_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_259_cast_fp16 = mul(x = clip_1_cast_fp16, y = clip_1_cast_fp16)[name = string("op_259_cast_fp16")];
            fp32 var_260_epsilon_0 = const()[name = string("op_260_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_260_cast_fp16 = inverse(epsilon = var_260_epsilon_0, x = var_259_cast_fp16)[name = string("op_260_cast_fp16")];
            fp16 var_261_to_fp16 = const()[name = string("op_261_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_262_cast_fp16 = mul(x = var_260_cast_fp16, y = var_261_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_263_cast_fp16 = add(x = var_3_cast_fp16, y = var_262_cast_fp16)[name = string("op_263_cast_fp16")];
            fp32 var_264_epsilon_0 = const()[name = string("op_264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_264_cast_fp16 = rsqrt(epsilon = var_264_epsilon_0, x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_3_cast_fp16 = mul(x = xn_3_cast_fp16, y = var_264_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [896]> lm_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138286784)))];
            tensor<fp16, [4, 64, 896]> input_3_cast_fp16 = mul(x = lm_layers_0_post_attention_layernorm_weight_to_fp16, y = out_3_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138288640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142646848))))[name = string("lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4864]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142656640)))];
            tensor<fp16, [4, 64, 4864]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_272_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_272_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142666432))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147024640))))[name = string("lm_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_7_cast_fp16 = mul(x = var_272_cast_fp16, y = linear_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147034432))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151392640))))[name = string("lm_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_6_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_3_cast_fp16 = add(x = x_1_cast_fp16, y = linear_6_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_284_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_284_cast_fp16")];
            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_286_cast_fp16 = reduce_max(axes = var_286_axes_0, keep_dims = var_12, x = var_284_cast_fp16)[name = string("op_286_cast_fp16")];
            fp16 const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_2_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_28_to_fp16, x = var_286_cast_fp16)[name = string("clip_2_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_5_cast_fp16 = real_div(x = x_3_cast_fp16, y = clip_2_cast_fp16)[name = string("xn_5_cast_fp16")];
            fp16 var_23_promoted_2_to_fp16 = const()[name = string("op_23_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_289_cast_fp16 = pow(x = xn_5_cast_fp16, y = var_23_promoted_2_to_fp16)[name = string("op_289_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_289_cast_fp16)[name = string("var_5_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_292_cast_fp16 = mul(x = clip_2_cast_fp16, y = clip_2_cast_fp16)[name = string("op_292_cast_fp16")];
            fp32 var_293_epsilon_0 = const()[name = string("op_293_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_293_cast_fp16 = inverse(epsilon = var_293_epsilon_0, x = var_292_cast_fp16)[name = string("op_293_cast_fp16")];
            fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_295_cast_fp16 = mul(x = var_293_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_296_cast_fp16 = add(x = var_5_cast_fp16, y = var_295_cast_fp16)[name = string("op_296_cast_fp16")];
            fp32 var_297_epsilon_0 = const()[name = string("op_297_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_297_cast_fp16 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16)[name = string("op_297_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_5_cast_fp16 = mul(x = xn_5_cast_fp16, y = var_297_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151394496)))];
            tensor<fp16, [4, 64, 896]> hidden_states_15_cast_fp16 = mul(x = lm_layers_1_input_layernorm_weight_to_fp16, y = out_5_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151396352))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152199232))))[name = string("lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152201088)))];
            tensor<fp16, [4, 64, 896]> linear_7_cast_fp16 = linear(bias = lm_layers_1_self_attn_q_proj_bias_to_fp16, weight = lm_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_309 = const()[name = string("op_309"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_310_cast_fp16 = reshape(shape = var_309, x = linear_7_cast_fp16)[name = string("op_310_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152202944))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152317696))))[name = string("lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318016)))];
            tensor<fp16, [4, 64, 128]> linear_8_cast_fp16 = linear(bias = lm_layers_1_self_attn_k_proj_bias_to_fp16, weight = lm_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_315 = const()[name = string("op_315"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_316_cast_fp16 = reshape(shape = var_315, x = linear_8_cast_fp16)[name = string("op_316_cast_fp16")];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152318336))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433088))))[name = string("lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433408)))];
            tensor<fp16, [4, 64, 128]> linear_9_cast_fp16 = linear(bias = lm_layers_1_self_attn_v_proj_bias_to_fp16, weight = lm_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_15_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_321 = const()[name = string("op_321"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_322_cast_fp16 = reshape(shape = var_321, x = linear_9_cast_fp16)[name = string("op_322_cast_fp16")];
            tensor<int32, [4]> hidden_states_21_perm_0 = const()[name = string("hidden_states_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_310_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [4, 14, 64, 64]> var_326_cast_fp16 = mul(x = q_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_3_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_3_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_337_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_337_cast_fp16")];
            bool var_339_interleave_0 = const()[name = string("op_339_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_339_cast_fp16 = concat(axis = var_20, interleave = var_339_interleave_0, values = (var_337_cast_fp16, x1_5_cast_fp16))[name = string("op_339_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_340_cast_fp16 = mul(x = var_339_cast_fp16, y = sin_5_cast_fp16)[name = string("op_340_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_340_cast_fp16)[name = string("query_3_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_316_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [4, 2, 64, 64]> var_342_cast_fp16 = mul(x = k_3_cast_fp16, y = cos_5_cast_fp16)[name = string("op_342_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_3_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_3_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_353_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_353_cast_fp16")];
            bool var_355_interleave_0 = const()[name = string("op_355_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_355_cast_fp16 = concat(axis = var_20, interleave = var_355_interleave_0, values = (var_353_cast_fp16, x1_7_cast_fp16))[name = string("op_355_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = sin_5_cast_fp16)[name = string("op_356_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_17_cast_fp16 = add(x = var_342_cast_fp16, y = var_356_cast_fp16)[name = string("hidden_states_17_cast_fp16")];
            tensor<int32, [1]> var_365_axes_0 = const()[name = string("op_365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_365_cast_fp16 = expand_dims(axes = var_365_axes_0, x = hidden_states_17_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<int32, [5]> hidden_states_19_reps_0 = const()[name = string("hidden_states_19_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_19_cast_fp16 = tile(reps = hidden_states_19_reps_0, x = var_365_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<int32, [4]> var_373 = const()[name = string("op_373"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_3_cast_fp16 = reshape(shape = var_373, x = hidden_states_19_cast_fp16)[name = string("key_3_cast_fp16")];
            tensor<int32, [1]> var_382_axes_0 = const()[name = string("op_382_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_21_cast_fp16 = transpose(perm = hidden_states_21_perm_0, x = var_322_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_382_cast_fp16 = expand_dims(axes = var_382_axes_0, x = hidden_states_21_cast_fp16)[name = string("op_382_cast_fp16")];
            tensor<int32, [5]> hidden_states_23_reps_0 = const()[name = string("hidden_states_23_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_23_cast_fp16 = tile(reps = hidden_states_23_reps_0, x = var_382_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_3_cast_fp16 = reshape(shape = var_390, x = hidden_states_23_cast_fp16)[name = string("value_3_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_3_cast_fp16 = mul(x = query_3_cast_fp16, y = var_13_to_fp16)[name = string("mul_3_cast_fp16")];
            bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)];
            bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = key_3_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_1_cast_fp16")];
            int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = add_1_cast_fp16)[name = string("softmax_1_cast_fp16")];
            bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)];
            bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = value_3_cast_fp16)[name = string("attn_output_5_cast_fp16")];
            tensor<int32, [4]> var_393_perm_0 = const()[name = string("op_393_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_395 = const()[name = string("op_395"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_393_cast_fp16 = transpose(perm = var_393_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [4, 64, 896]> var_396_cast_fp16 = reshape(shape = var_395, x = var_393_cast_fp16)[name = string("op_396_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152433728))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153236608))))[name = string("lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_10_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = var_396_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = linear_10_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_402_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_402_cast_fp16")];
            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_404_cast_fp16 = reduce_max(axes = var_404_axes_0, keep_dims = var_12, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")];
            fp16 const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_3_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_45_to_fp16, x = var_404_cast_fp16)[name = string("clip_3_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = clip_3_cast_fp16)[name = string("xn_7_cast_fp16")];
            fp16 var_23_promoted_3_to_fp16 = const()[name = string("op_23_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_407_cast_fp16 = pow(x = xn_7_cast_fp16, y = var_23_promoted_3_to_fp16)[name = string("op_407_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_407_cast_fp16)[name = string("var_7_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_410_cast_fp16 = mul(x = clip_3_cast_fp16, y = clip_3_cast_fp16)[name = string("op_410_cast_fp16")];
            fp32 var_411_epsilon_0 = const()[name = string("op_411_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_411_cast_fp16 = inverse(epsilon = var_411_epsilon_0, x = var_410_cast_fp16)[name = string("op_411_cast_fp16")];
            fp16 var_412_to_fp16 = const()[name = string("op_412_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_413_cast_fp16 = mul(x = var_411_cast_fp16, y = var_412_to_fp16)[name = string("op_413_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_414_cast_fp16 = add(x = var_7_cast_fp16, y = var_413_cast_fp16)[name = string("op_414_cast_fp16")];
            fp32 var_415_epsilon_0 = const()[name = string("op_415_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_415_cast_fp16 = rsqrt(epsilon = var_415_epsilon_0, x = var_414_cast_fp16)[name = string("op_415_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_7_cast_fp16 = mul(x = xn_7_cast_fp16, y = var_415_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [896]> lm_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153238464)))];
            tensor<fp16, [4, 64, 896]> input_11_cast_fp16 = mul(x = lm_layers_1_post_attention_layernorm_weight_to_fp16, y = out_7_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153240320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157598528))))[name = string("lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_423_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_423_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157608320))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161966528))))[name = string("lm_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_15_cast_fp16 = mul(x = var_423_cast_fp16, y = linear_12_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161976320))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166334528))))[name = string("lm_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_13_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_13_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_435_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_435_cast_fp16")];
            tensor<int32, [1]> var_437_axes_0 = const()[name = string("op_437_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_437_cast_fp16 = reduce_max(axes = var_437_axes_0, keep_dims = var_12, x = var_435_cast_fp16)[name = string("op_437_cast_fp16")];
            fp16 const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_4_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_46_to_fp16, x = var_437_cast_fp16)[name = string("clip_4_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_9_cast_fp16 = real_div(x = x_7_cast_fp16, y = clip_4_cast_fp16)[name = string("xn_9_cast_fp16")];
            fp16 var_23_promoted_4_to_fp16 = const()[name = string("op_23_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_440_cast_fp16 = pow(x = xn_9_cast_fp16, y = var_23_promoted_4_to_fp16)[name = string("op_440_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_440_cast_fp16)[name = string("var_9_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_443_cast_fp16 = mul(x = clip_4_cast_fp16, y = clip_4_cast_fp16)[name = string("op_443_cast_fp16")];
            fp32 var_444_epsilon_0 = const()[name = string("op_444_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_444_cast_fp16 = inverse(epsilon = var_444_epsilon_0, x = var_443_cast_fp16)[name = string("op_444_cast_fp16")];
            fp16 var_445_to_fp16 = const()[name = string("op_445_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_446_cast_fp16 = mul(x = var_444_cast_fp16, y = var_445_to_fp16)[name = string("op_446_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_447_cast_fp16 = add(x = var_9_cast_fp16, y = var_446_cast_fp16)[name = string("op_447_cast_fp16")];
            fp32 var_448_epsilon_0 = const()[name = string("op_448_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_448_cast_fp16 = rsqrt(epsilon = var_448_epsilon_0, x = var_447_cast_fp16)[name = string("op_448_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_9_cast_fp16 = mul(x = xn_9_cast_fp16, y = var_448_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166336384)))];
            tensor<fp16, [4, 64, 896]> hidden_states_29_cast_fp16 = mul(x = lm_layers_2_input_layernorm_weight_to_fp16, y = out_9_cast_fp16)[name = string("hidden_states_29_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166338240))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167141120))))[name = string("lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167142976)))];
            tensor<fp16, [4, 64, 896]> linear_14_cast_fp16 = linear(bias = lm_layers_2_self_attn_q_proj_bias_to_fp16, weight = lm_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_14_cast_fp16)[name = string("op_461_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = string("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167144832))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259584))))[name = string("lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167259904)))];
            tensor<fp16, [4, 64, 128]> linear_15_cast_fp16 = linear(bias = lm_layers_2_self_attn_k_proj_bias_to_fp16, weight = lm_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_15_cast_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167260224))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167374976))))[name = string("lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375296)))];
            tensor<fp16, [4, 64, 128]> linear_16_cast_fp16 = linear(bias = lm_layers_2_self_attn_v_proj_bias_to_fp16, weight = lm_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_29_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_16_cast_fp16)[name = string("op_473_cast_fp16")];
            tensor<int32, [4]> hidden_states_35_perm_0 = const()[name = string("hidden_states_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_461_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [4, 14, 64, 64]> var_477_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_477_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_5_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_5_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_488_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_488_cast_fp16")];
            bool var_490_interleave_0 = const()[name = string("op_490_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_490_cast_fp16 = concat(axis = var_20, interleave = var_490_interleave_0, values = (var_488_cast_fp16, x1_9_cast_fp16))[name = string("op_490_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_491_cast_fp16 = mul(x = var_490_cast_fp16, y = sin_5_cast_fp16)[name = string("op_491_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_5_cast_fp16 = add(x = var_477_cast_fp16, y = var_491_cast_fp16)[name = string("query_5_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_467_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [4, 2, 64, 64]> var_493_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_5_cast_fp16)[name = string("op_493_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_5_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_5_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_504_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_504_cast_fp16")];
            bool var_506_interleave_0 = const()[name = string("op_506_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_506_cast_fp16 = concat(axis = var_20, interleave = var_506_interleave_0, values = (var_504_cast_fp16, x1_11_cast_fp16))[name = string("op_506_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_507_cast_fp16 = mul(x = var_506_cast_fp16, y = sin_5_cast_fp16)[name = string("op_507_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_31_cast_fp16 = add(x = var_493_cast_fp16, y = var_507_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [1]> var_516_axes_0 = const()[name = string("op_516_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_516_cast_fp16 = expand_dims(axes = var_516_axes_0, x = hidden_states_31_cast_fp16)[name = string("op_516_cast_fp16")];
            tensor<int32, [5]> hidden_states_33_reps_0 = const()[name = string("hidden_states_33_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_33_cast_fp16 = tile(reps = hidden_states_33_reps_0, x = var_516_cast_fp16)[name = string("hidden_states_33_cast_fp16")];
            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_5_cast_fp16 = reshape(shape = var_524, x = hidden_states_33_cast_fp16)[name = string("key_5_cast_fp16")];
            tensor<int32, [1]> var_533_axes_0 = const()[name = string("op_533_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_35_cast_fp16 = transpose(perm = hidden_states_35_perm_0, x = var_473_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_533_cast_fp16 = expand_dims(axes = var_533_axes_0, x = hidden_states_35_cast_fp16)[name = string("op_533_cast_fp16")];
            tensor<int32, [5]> hidden_states_37_reps_0 = const()[name = string("hidden_states_37_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_37_cast_fp16 = tile(reps = hidden_states_37_reps_0, x = var_533_cast_fp16)[name = string("hidden_states_37_cast_fp16")];
            tensor<int32, [4]> var_541 = const()[name = string("op_541"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_5_cast_fp16 = reshape(shape = var_541, x = hidden_states_37_cast_fp16)[name = string("value_5_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_5_cast_fp16 = mul(x = query_5_cast_fp16, y = var_13_to_fp16)[name = string("mul_5_cast_fp16")];
            bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)];
            bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = key_5_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_2_cast_fp16")];
            int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = add_2_cast_fp16)[name = string("softmax_2_cast_fp16")];
            bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)];
            bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = value_5_cast_fp16)[name = string("attn_output_9_cast_fp16")];
            tensor<int32, [4]> var_544_perm_0 = const()[name = string("op_544_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_546 = const()[name = string("op_546"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_544_cast_fp16 = transpose(perm = var_544_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [4, 64, 896]> var_547_cast_fp16 = reshape(shape = var_546, x = var_544_cast_fp16)[name = string("op_547_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167375616))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168178496))))[name = string("lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_17_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = var_547_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = linear_17_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_553_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_553_cast_fp16")];
            tensor<int32, [1]> var_555_axes_0 = const()[name = string("op_555_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_555_cast_fp16 = reduce_max(axes = var_555_axes_0, keep_dims = var_12, x = var_553_cast_fp16)[name = string("op_555_cast_fp16")];
            fp16 const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_5_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_63_to_fp16, x = var_555_cast_fp16)[name = string("clip_5_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = clip_5_cast_fp16)[name = string("xn_11_cast_fp16")];
            fp16 var_23_promoted_5_to_fp16 = const()[name = string("op_23_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_558_cast_fp16 = pow(x = xn_11_cast_fp16, y = var_23_promoted_5_to_fp16)[name = string("op_558_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_558_cast_fp16)[name = string("var_11_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_561_cast_fp16 = mul(x = clip_5_cast_fp16, y = clip_5_cast_fp16)[name = string("op_561_cast_fp16")];
            fp32 var_562_epsilon_0 = const()[name = string("op_562_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_562_cast_fp16 = inverse(epsilon = var_562_epsilon_0, x = var_561_cast_fp16)[name = string("op_562_cast_fp16")];
            fp16 var_563_to_fp16 = const()[name = string("op_563_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_564_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_565_cast_fp16 = add(x = var_11_cast_fp16, y = var_564_cast_fp16)[name = string("op_565_cast_fp16")];
            fp32 var_566_epsilon_0 = const()[name = string("op_566_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_566_cast_fp16 = rsqrt(epsilon = var_566_epsilon_0, x = var_565_cast_fp16)[name = string("op_566_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_11_cast_fp16 = mul(x = xn_11_cast_fp16, y = var_566_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [896]> lm_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168180352)))];
            tensor<fp16, [4, 64, 896]> input_19_cast_fp16 = mul(x = lm_layers_2_post_attention_layernorm_weight_to_fp16, y = out_11_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168182208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172540416))))[name = string("lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_574_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172550208))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176908416))))[name = string("lm_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_23_cast_fp16 = mul(x = var_574_cast_fp16, y = linear_19_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176918208))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181276416))))[name = string("lm_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_20_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_11_cast_fp16 = add(x = x_9_cast_fp16, y = linear_20_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_586_cast_fp16 = abs(x = x_11_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_588_axes_0 = const()[name = string("op_588_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_588_cast_fp16 = reduce_max(axes = var_588_axes_0, keep_dims = var_12, x = var_586_cast_fp16)[name = string("op_588_cast_fp16")];
            fp16 const_64_to_fp16 = const()[name = string("const_64_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_6_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_64_to_fp16, x = var_588_cast_fp16)[name = string("clip_6_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_13_cast_fp16 = real_div(x = x_11_cast_fp16, y = clip_6_cast_fp16)[name = string("xn_13_cast_fp16")];
            fp16 var_23_promoted_6_to_fp16 = const()[name = string("op_23_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_591_cast_fp16 = pow(x = xn_13_cast_fp16, y = var_23_promoted_6_to_fp16)[name = string("op_591_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_591_cast_fp16)[name = string("var_13_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_594_cast_fp16 = mul(x = clip_6_cast_fp16, y = clip_6_cast_fp16)[name = string("op_594_cast_fp16")];
            fp32 var_595_epsilon_0 = const()[name = string("op_595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_595_cast_fp16 = inverse(epsilon = var_595_epsilon_0, x = var_594_cast_fp16)[name = string("op_595_cast_fp16")];
            fp16 var_596_to_fp16 = const()[name = string("op_596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_597_cast_fp16 = mul(x = var_595_cast_fp16, y = var_596_to_fp16)[name = string("op_597_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_598_cast_fp16 = add(x = var_13_cast_fp16, y = var_597_cast_fp16)[name = string("op_598_cast_fp16")];
            fp32 var_599_epsilon_0 = const()[name = string("op_599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_599_cast_fp16 = rsqrt(epsilon = var_599_epsilon_0, x = var_598_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_13_cast_fp16 = mul(x = xn_13_cast_fp16, y = var_599_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181278272)))];
            tensor<fp16, [4, 64, 896]> hidden_states_43_cast_fp16 = mul(x = lm_layers_3_input_layernorm_weight_to_fp16, y = out_13_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(181280128))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182083008))))[name = string("lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182084864)))];
            tensor<fp16, [4, 64, 896]> linear_21_cast_fp16 = linear(bias = lm_layers_3_self_attn_q_proj_bias_to_fp16, weight = lm_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_611 = const()[name = string("op_611"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_612_cast_fp16 = reshape(shape = var_611, x = linear_21_cast_fp16)[name = string("op_612_cast_fp16")];
            tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182086720))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201472))))[name = string("lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182201792)))];
            tensor<fp16, [4, 64, 128]> linear_22_cast_fp16 = linear(bias = lm_layers_3_self_attn_k_proj_bias_to_fp16, weight = lm_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_617 = const()[name = string("op_617"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_618_cast_fp16 = reshape(shape = var_617, x = linear_22_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182202112))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182316864))))[name = string("lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317184)))];
            tensor<fp16, [4, 64, 128]> linear_23_cast_fp16 = linear(bias = lm_layers_3_self_attn_v_proj_bias_to_fp16, weight = lm_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_43_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_623 = const()[name = string("op_623"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_624_cast_fp16 = reshape(shape = var_623, x = linear_23_cast_fp16)[name = string("op_624_cast_fp16")];
            tensor<int32, [4]> hidden_states_49_perm_0 = const()[name = string("hidden_states_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_7_cast_fp16 = transpose(perm = q_7_perm_0, x = var_612_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [4, 14, 64, 64]> var_628_cast_fp16 = mul(x = q_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_628_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_7_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_7_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_69_promoted_to_fp16 = const()[name = string("const_69_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_639_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_69_promoted_to_fp16)[name = string("op_639_cast_fp16")];
            bool var_641_interleave_0 = const()[name = string("op_641_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_641_cast_fp16 = concat(axis = var_20, interleave = var_641_interleave_0, values = (var_639_cast_fp16, x1_13_cast_fp16))[name = string("op_641_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_642_cast_fp16 = mul(x = var_641_cast_fp16, y = sin_5_cast_fp16)[name = string("op_642_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_7_cast_fp16 = add(x = var_628_cast_fp16, y = var_642_cast_fp16)[name = string("query_7_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = var_618_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [4, 2, 64, 64]> var_644_cast_fp16 = mul(x = k_7_cast_fp16, y = cos_5_cast_fp16)[name = string("op_644_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_7_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_7_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_72_promoted_to_fp16 = const()[name = string("const_72_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_655_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_72_promoted_to_fp16)[name = string("op_655_cast_fp16")];
            bool var_657_interleave_0 = const()[name = string("op_657_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_657_cast_fp16 = concat(axis = var_20, interleave = var_657_interleave_0, values = (var_655_cast_fp16, x1_15_cast_fp16))[name = string("op_657_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_658_cast_fp16 = mul(x = var_657_cast_fp16, y = sin_5_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_45_cast_fp16 = add(x = var_644_cast_fp16, y = var_658_cast_fp16)[name = string("hidden_states_45_cast_fp16")];
            tensor<int32, [1]> var_667_axes_0 = const()[name = string("op_667_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_667_cast_fp16 = expand_dims(axes = var_667_axes_0, x = hidden_states_45_cast_fp16)[name = string("op_667_cast_fp16")];
            tensor<int32, [5]> hidden_states_47_reps_0 = const()[name = string("hidden_states_47_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_47_cast_fp16 = tile(reps = hidden_states_47_reps_0, x = var_667_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> var_675 = const()[name = string("op_675"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_7_cast_fp16 = reshape(shape = var_675, x = hidden_states_47_cast_fp16)[name = string("key_7_cast_fp16")];
            tensor<int32, [1]> var_684_axes_0 = const()[name = string("op_684_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_49_cast_fp16 = transpose(perm = hidden_states_49_perm_0, x = var_624_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_684_cast_fp16 = expand_dims(axes = var_684_axes_0, x = hidden_states_49_cast_fp16)[name = string("op_684_cast_fp16")];
            tensor<int32, [5]> hidden_states_51_reps_0 = const()[name = string("hidden_states_51_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_51_cast_fp16 = tile(reps = hidden_states_51_reps_0, x = var_684_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<int32, [4]> var_692 = const()[name = string("op_692"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_7_cast_fp16 = reshape(shape = var_692, x = hidden_states_51_cast_fp16)[name = string("value_7_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_7_cast_fp16 = mul(x = query_7_cast_fp16, y = var_13_to_fp16)[name = string("mul_7_cast_fp16")];
            bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)];
            bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = key_7_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_3_cast_fp16")];
            int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = add_3_cast_fp16)[name = string("softmax_3_cast_fp16")];
            bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)];
            bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = value_7_cast_fp16)[name = string("attn_output_13_cast_fp16")];
            tensor<int32, [4]> var_695_perm_0 = const()[name = string("op_695_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_697 = const()[name = string("op_697"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_695_cast_fp16 = transpose(perm = var_695_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [4, 64, 896]> var_698_cast_fp16 = reshape(shape = var_697, x = var_695_cast_fp16)[name = string("op_698_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182317504))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183120384))))[name = string("lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_24_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = var_698_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_24_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_704_cast_fp16 = abs(x = x_13_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_706_axes_0 = const()[name = string("op_706_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_706_cast_fp16 = reduce_max(axes = var_706_axes_0, keep_dims = var_12, x = var_704_cast_fp16)[name = string("op_706_cast_fp16")];
            fp16 const_81_to_fp16 = const()[name = string("const_81_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_7_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_81_to_fp16, x = var_706_cast_fp16)[name = string("clip_7_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_15_cast_fp16 = real_div(x = x_13_cast_fp16, y = clip_7_cast_fp16)[name = string("xn_15_cast_fp16")];
            fp16 var_23_promoted_7_to_fp16 = const()[name = string("op_23_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_709_cast_fp16 = pow(x = xn_15_cast_fp16, y = var_23_promoted_7_to_fp16)[name = string("op_709_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_709_cast_fp16)[name = string("var_15_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_712_cast_fp16 = mul(x = clip_7_cast_fp16, y = clip_7_cast_fp16)[name = string("op_712_cast_fp16")];
            fp32 var_713_epsilon_0 = const()[name = string("op_713_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_713_cast_fp16 = inverse(epsilon = var_713_epsilon_0, x = var_712_cast_fp16)[name = string("op_713_cast_fp16")];
            fp16 var_714_to_fp16 = const()[name = string("op_714_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_715_cast_fp16 = mul(x = var_713_cast_fp16, y = var_714_to_fp16)[name = string("op_715_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_716_cast_fp16 = add(x = var_15_cast_fp16, y = var_715_cast_fp16)[name = string("op_716_cast_fp16")];
            fp32 var_717_epsilon_0 = const()[name = string("op_717_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_717_cast_fp16 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_15_cast_fp16 = mul(x = xn_15_cast_fp16, y = var_717_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [896]> lm_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183122240)))];
            tensor<fp16, [4, 64, 896]> input_27_cast_fp16 = mul(x = lm_layers_3_post_attention_layernorm_weight_to_fp16, y = out_15_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183124096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187482304))))[name = string("lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_725_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187492096))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191850304))))[name = string("lm_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_31_cast_fp16 = mul(x = var_725_cast_fp16, y = linear_26_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(191860096))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218304))))[name = string("lm_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_15_cast_fp16 = add(x = x_13_cast_fp16, y = linear_27_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_737_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_737_cast_fp16")];
            tensor<int32, [1]> var_739_axes_0 = const()[name = string("op_739_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_739_cast_fp16 = reduce_max(axes = var_739_axes_0, keep_dims = var_12, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 const_82_to_fp16 = const()[name = string("const_82_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_8_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_82_to_fp16, x = var_739_cast_fp16)[name = string("clip_8_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_17_cast_fp16 = real_div(x = x_15_cast_fp16, y = clip_8_cast_fp16)[name = string("xn_17_cast_fp16")];
            fp16 var_23_promoted_8_to_fp16 = const()[name = string("op_23_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_742_cast_fp16 = pow(x = xn_17_cast_fp16, y = var_23_promoted_8_to_fp16)[name = string("op_742_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_742_cast_fp16)[name = string("var_17_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_745_cast_fp16 = mul(x = clip_8_cast_fp16, y = clip_8_cast_fp16)[name = string("op_745_cast_fp16")];
            fp32 var_746_epsilon_0 = const()[name = string("op_746_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_746_cast_fp16 = inverse(epsilon = var_746_epsilon_0, x = var_745_cast_fp16)[name = string("op_746_cast_fp16")];
            fp16 var_747_to_fp16 = const()[name = string("op_747_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_748_cast_fp16 = mul(x = var_746_cast_fp16, y = var_747_to_fp16)[name = string("op_748_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_749_cast_fp16 = add(x = var_17_cast_fp16, y = var_748_cast_fp16)[name = string("op_749_cast_fp16")];
            fp32 var_750_epsilon_0 = const()[name = string("op_750_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_750_cast_fp16 = rsqrt(epsilon = var_750_epsilon_0, x = var_749_cast_fp16)[name = string("op_750_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_17_cast_fp16 = mul(x = xn_17_cast_fp16, y = var_750_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220160)))];
            tensor<fp16, [4, 64, 896]> hidden_states_57_cast_fp16 = mul(x = lm_layers_4_input_layernorm_weight_to_fp16, y = out_17_cast_fp16)[name = string("hidden_states_57_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196222016))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197024896))))[name = string("lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197026752)))];
            tensor<fp16, [4, 64, 896]> linear_28_cast_fp16 = linear(bias = lm_layers_4_self_attn_q_proj_bias_to_fp16, weight = lm_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_763_cast_fp16 = reshape(shape = var_762, x = linear_28_cast_fp16)[name = string("op_763_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197028608))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143360))))[name = string("lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197143680)))];
            tensor<fp16, [4, 64, 128]> linear_29_cast_fp16 = linear(bias = lm_layers_4_self_attn_k_proj_bias_to_fp16, weight = lm_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_768 = const()[name = string("op_768"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_769_cast_fp16 = reshape(shape = var_768, x = linear_29_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = string("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197144000))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197258752))))[name = string("lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259072)))];
            tensor<fp16, [4, 64, 128]> linear_30_cast_fp16 = linear(bias = lm_layers_4_self_attn_v_proj_bias_to_fp16, weight = lm_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_57_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_775_cast_fp16 = reshape(shape = var_774, x = linear_30_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<int32, [4]> hidden_states_63_perm_0 = const()[name = string("hidden_states_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_763_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [4, 14, 64, 64]> var_779_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_9_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_9_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_87_promoted_to_fp16 = const()[name = string("const_87_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_790_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_87_promoted_to_fp16)[name = string("op_790_cast_fp16")];
            bool var_792_interleave_0 = const()[name = string("op_792_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_792_cast_fp16 = concat(axis = var_20, interleave = var_792_interleave_0, values = (var_790_cast_fp16, x1_17_cast_fp16))[name = string("op_792_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_793_cast_fp16 = mul(x = var_792_cast_fp16, y = sin_5_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_793_cast_fp16)[name = string("query_9_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_769_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [4, 2, 64, 64]> var_795_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_5_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_9_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_9_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_90_promoted_to_fp16 = const()[name = string("const_90_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_806_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_90_promoted_to_fp16)[name = string("op_806_cast_fp16")];
            bool var_808_interleave_0 = const()[name = string("op_808_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_808_cast_fp16 = concat(axis = var_20, interleave = var_808_interleave_0, values = (var_806_cast_fp16, x1_19_cast_fp16))[name = string("op_808_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_809_cast_fp16 = mul(x = var_808_cast_fp16, y = sin_5_cast_fp16)[name = string("op_809_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_59_cast_fp16 = add(x = var_795_cast_fp16, y = var_809_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<int32, [1]> var_818_axes_0 = const()[name = string("op_818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_818_cast_fp16 = expand_dims(axes = var_818_axes_0, x = hidden_states_59_cast_fp16)[name = string("op_818_cast_fp16")];
            tensor<int32, [5]> hidden_states_61_reps_0 = const()[name = string("hidden_states_61_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_61_cast_fp16 = tile(reps = hidden_states_61_reps_0, x = var_818_cast_fp16)[name = string("hidden_states_61_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_9_cast_fp16 = reshape(shape = var_826, x = hidden_states_61_cast_fp16)[name = string("key_9_cast_fp16")];
            tensor<int32, [1]> var_835_axes_0 = const()[name = string("op_835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_63_cast_fp16 = transpose(perm = hidden_states_63_perm_0, x = var_775_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_835_cast_fp16 = expand_dims(axes = var_835_axes_0, x = hidden_states_63_cast_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [5]> hidden_states_65_reps_0 = const()[name = string("hidden_states_65_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_65_cast_fp16 = tile(reps = hidden_states_65_reps_0, x = var_835_cast_fp16)[name = string("hidden_states_65_cast_fp16")];
            tensor<int32, [4]> var_843 = const()[name = string("op_843"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_9_cast_fp16 = reshape(shape = var_843, x = hidden_states_65_cast_fp16)[name = string("value_9_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_9_cast_fp16 = mul(x = query_9_cast_fp16, y = var_13_to_fp16)[name = string("mul_9_cast_fp16")];
            bool matmul_4_transpose_y_0 = const()[name = string("matmul_4_transpose_y_0"), val = bool(true)];
            bool matmul_4_transpose_x_0 = const()[name = string("matmul_4_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = key_9_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_4_cast_fp16")];
            int32 softmax_4_axis_0 = const()[name = string("softmax_4_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = add_4_cast_fp16)[name = string("softmax_4_cast_fp16")];
            bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)];
            bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_4_cast_fp16, y = value_9_cast_fp16)[name = string("attn_output_17_cast_fp16")];
            tensor<int32, [4]> var_846_perm_0 = const()[name = string("op_846_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_848 = const()[name = string("op_848"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_846_cast_fp16 = transpose(perm = var_846_perm_0, x = attn_output_17_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [4, 64, 896]> var_849_cast_fp16 = reshape(shape = var_848, x = var_846_cast_fp16)[name = string("op_849_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197259392))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198062272))))[name = string("lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = var_849_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = linear_31_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_855_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_855_cast_fp16")];
            tensor<int32, [1]> var_857_axes_0 = const()[name = string("op_857_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_857_cast_fp16 = reduce_max(axes = var_857_axes_0, keep_dims = var_12, x = var_855_cast_fp16)[name = string("op_857_cast_fp16")];
            fp16 const_99_to_fp16 = const()[name = string("const_99_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_9_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_99_to_fp16, x = var_857_cast_fp16)[name = string("clip_9_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = clip_9_cast_fp16)[name = string("xn_19_cast_fp16")];
            fp16 var_23_promoted_9_to_fp16 = const()[name = string("op_23_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_860_cast_fp16 = pow(x = xn_19_cast_fp16, y = var_23_promoted_9_to_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_860_cast_fp16)[name = string("var_19_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_863_cast_fp16 = mul(x = clip_9_cast_fp16, y = clip_9_cast_fp16)[name = string("op_863_cast_fp16")];
            fp32 var_864_epsilon_0 = const()[name = string("op_864_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_864_cast_fp16 = inverse(epsilon = var_864_epsilon_0, x = var_863_cast_fp16)[name = string("op_864_cast_fp16")];
            fp16 var_865_to_fp16 = const()[name = string("op_865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_866_cast_fp16 = mul(x = var_864_cast_fp16, y = var_865_to_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_867_cast_fp16 = add(x = var_19_cast_fp16, y = var_866_cast_fp16)[name = string("op_867_cast_fp16")];
            fp32 var_868_epsilon_0 = const()[name = string("op_868_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_868_cast_fp16 = rsqrt(epsilon = var_868_epsilon_0, x = var_867_cast_fp16)[name = string("op_868_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_19_cast_fp16 = mul(x = xn_19_cast_fp16, y = var_868_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [896]> lm_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198064128)))];
            tensor<fp16, [4, 64, 896]> input_35_cast_fp16 = mul(x = lm_layers_4_post_attention_layernorm_weight_to_fp16, y = out_19_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198065984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202424192))))[name = string("lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_876_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_876_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202433984))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206792192))))[name = string("lm_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_39_cast_fp16 = mul(x = var_876_cast_fp16, y = linear_33_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206801984))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211160192))))[name = string("lm_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_34_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_34_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_888_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<int32, [1]> var_890_axes_0 = const()[name = string("op_890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_890_cast_fp16 = reduce_max(axes = var_890_axes_0, keep_dims = var_12, x = var_888_cast_fp16)[name = string("op_890_cast_fp16")];
            fp16 const_100_to_fp16 = const()[name = string("const_100_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_10_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_100_to_fp16, x = var_890_cast_fp16)[name = string("clip_10_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_21_cast_fp16 = real_div(x = x_19_cast_fp16, y = clip_10_cast_fp16)[name = string("xn_21_cast_fp16")];
            fp16 var_23_promoted_10_to_fp16 = const()[name = string("op_23_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_893_cast_fp16 = pow(x = xn_21_cast_fp16, y = var_23_promoted_10_to_fp16)[name = string("op_893_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_893_cast_fp16)[name = string("var_21_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_896_cast_fp16 = mul(x = clip_10_cast_fp16, y = clip_10_cast_fp16)[name = string("op_896_cast_fp16")];
            fp32 var_897_epsilon_0 = const()[name = string("op_897_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_897_cast_fp16 = inverse(epsilon = var_897_epsilon_0, x = var_896_cast_fp16)[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_900_cast_fp16 = add(x = var_21_cast_fp16, y = var_899_cast_fp16)[name = string("op_900_cast_fp16")];
            fp32 var_901_epsilon_0 = const()[name = string("op_901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_901_cast_fp16 = rsqrt(epsilon = var_901_epsilon_0, x = var_900_cast_fp16)[name = string("op_901_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_21_cast_fp16 = mul(x = xn_21_cast_fp16, y = var_901_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211162048)))];
            tensor<fp16, [4, 64, 896]> hidden_states_71_cast_fp16 = mul(x = lm_layers_5_input_layernorm_weight_to_fp16, y = out_21_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211163904))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211966784))))[name = string("lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211968640)))];
            tensor<fp16, [4, 64, 896]> linear_35_cast_fp16 = linear(bias = lm_layers_5_self_attn_q_proj_bias_to_fp16, weight = lm_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_913 = const()[name = string("op_913"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_914_cast_fp16 = reshape(shape = var_913, x = linear_35_cast_fp16)[name = string("op_914_cast_fp16")];
            tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211970496))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085248))))[name = string("lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085568)))];
            tensor<fp16, [4, 64, 128]> linear_36_cast_fp16 = linear(bias = lm_layers_5_self_attn_k_proj_bias_to_fp16, weight = lm_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_920_cast_fp16 = reshape(shape = var_919, x = linear_36_cast_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212085888))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200640))))[name = string("lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212200960)))];
            tensor<fp16, [4, 64, 128]> linear_37_cast_fp16 = linear(bias = lm_layers_5_self_attn_v_proj_bias_to_fp16, weight = lm_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_71_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_926_cast_fp16 = reshape(shape = var_925, x = linear_37_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<int32, [4]> hidden_states_77_perm_0 = const()[name = string("hidden_states_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_11_cast_fp16 = transpose(perm = q_11_perm_0, x = var_914_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [4, 14, 64, 64]> var_930_cast_fp16 = mul(x = q_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_930_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_11_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_11_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_105_promoted_to_fp16 = const()[name = string("const_105_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_941_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_105_promoted_to_fp16)[name = string("op_941_cast_fp16")];
            bool var_943_interleave_0 = const()[name = string("op_943_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_943_cast_fp16 = concat(axis = var_20, interleave = var_943_interleave_0, values = (var_941_cast_fp16, x1_21_cast_fp16))[name = string("op_943_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_944_cast_fp16 = mul(x = var_943_cast_fp16, y = sin_5_cast_fp16)[name = string("op_944_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_11_cast_fp16 = add(x = var_930_cast_fp16, y = var_944_cast_fp16)[name = string("query_11_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = var_920_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [4, 2, 64, 64]> var_946_cast_fp16 = mul(x = k_11_cast_fp16, y = cos_5_cast_fp16)[name = string("op_946_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_11_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_11_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_108_promoted_to_fp16 = const()[name = string("const_108_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_957_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_108_promoted_to_fp16)[name = string("op_957_cast_fp16")];
            bool var_959_interleave_0 = const()[name = string("op_959_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_959_cast_fp16 = concat(axis = var_20, interleave = var_959_interleave_0, values = (var_957_cast_fp16, x1_23_cast_fp16))[name = string("op_959_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_960_cast_fp16 = mul(x = var_959_cast_fp16, y = sin_5_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_73_cast_fp16 = add(x = var_946_cast_fp16, y = var_960_cast_fp16)[name = string("hidden_states_73_cast_fp16")];
            tensor<int32, [1]> var_969_axes_0 = const()[name = string("op_969_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_969_cast_fp16 = expand_dims(axes = var_969_axes_0, x = hidden_states_73_cast_fp16)[name = string("op_969_cast_fp16")];
            tensor<int32, [5]> hidden_states_75_reps_0 = const()[name = string("hidden_states_75_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_75_cast_fp16 = tile(reps = hidden_states_75_reps_0, x = var_969_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<int32, [4]> var_977 = const()[name = string("op_977"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_11_cast_fp16 = reshape(shape = var_977, x = hidden_states_75_cast_fp16)[name = string("key_11_cast_fp16")];
            tensor<int32, [1]> var_986_axes_0 = const()[name = string("op_986_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_77_cast_fp16 = transpose(perm = hidden_states_77_perm_0, x = var_926_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_986_cast_fp16 = expand_dims(axes = var_986_axes_0, x = hidden_states_77_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<int32, [5]> hidden_states_79_reps_0 = const()[name = string("hidden_states_79_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_79_cast_fp16 = tile(reps = hidden_states_79_reps_0, x = var_986_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> var_994 = const()[name = string("op_994"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_11_cast_fp16 = reshape(shape = var_994, x = hidden_states_79_cast_fp16)[name = string("value_11_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_11_cast_fp16 = mul(x = query_11_cast_fp16, y = var_13_to_fp16)[name = string("mul_11_cast_fp16")];
            bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)];
            bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_5_cast_fp16")];
            int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = add_5_cast_fp16)[name = string("softmax_5_cast_fp16")];
            bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)];
            bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")];
            tensor<int32, [4]> var_997_perm_0 = const()[name = string("op_997_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_999 = const()[name = string("op_999"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_997_cast_fp16 = transpose(perm = var_997_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [4, 64, 896]> var_1000_cast_fp16 = reshape(shape = var_999, x = var_997_cast_fp16)[name = string("op_1000_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212201280))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213004160))))[name = string("lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_38_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = var_1000_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_21_cast_fp16 = add(x = x_19_cast_fp16, y = linear_38_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1006_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1006_cast_fp16")];
            tensor<int32, [1]> var_1008_axes_0 = const()[name = string("op_1008_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1008_cast_fp16 = reduce_max(axes = var_1008_axes_0, keep_dims = var_12, x = var_1006_cast_fp16)[name = string("op_1008_cast_fp16")];
            fp16 const_117_to_fp16 = const()[name = string("const_117_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_11_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_117_to_fp16, x = var_1008_cast_fp16)[name = string("clip_11_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = clip_11_cast_fp16)[name = string("xn_23_cast_fp16")];
            fp16 var_23_promoted_11_to_fp16 = const()[name = string("op_23_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1011_cast_fp16 = pow(x = xn_23_cast_fp16, y = var_23_promoted_11_to_fp16)[name = string("op_1011_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_1011_cast_fp16)[name = string("var_23_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1014_cast_fp16 = mul(x = clip_11_cast_fp16, y = clip_11_cast_fp16)[name = string("op_1014_cast_fp16")];
            fp32 var_1015_epsilon_0 = const()[name = string("op_1015_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1015_cast_fp16 = inverse(epsilon = var_1015_epsilon_0, x = var_1014_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1018_cast_fp16 = add(x = var_23_cast_fp16, y = var_1017_cast_fp16)[name = string("op_1018_cast_fp16")];
            fp32 var_1019_epsilon_0 = const()[name = string("op_1019_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1019_cast_fp16 = rsqrt(epsilon = var_1019_epsilon_0, x = var_1018_cast_fp16)[name = string("op_1019_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_23_cast_fp16 = mul(x = xn_23_cast_fp16, y = var_1019_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [896]> lm_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213006016)))];
            tensor<fp16, [4, 64, 896]> input_43_cast_fp16 = mul(x = lm_layers_5_post_attention_layernorm_weight_to_fp16, y = out_23_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213007872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217366080))))[name = string("lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1027_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1027_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217375872))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221734080))))[name = string("lm_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_47_cast_fp16 = mul(x = var_1027_cast_fp16, y = linear_40_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221743872))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226102080))))[name = string("lm_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_41_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_23_cast_fp16 = add(x = x_21_cast_fp16, y = linear_41_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1039_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1039_cast_fp16")];
            tensor<int32, [1]> var_1041_axes_0 = const()[name = string("op_1041_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1041_cast_fp16 = reduce_max(axes = var_1041_axes_0, keep_dims = var_12, x = var_1039_cast_fp16)[name = string("op_1041_cast_fp16")];
            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_12_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_118_to_fp16, x = var_1041_cast_fp16)[name = string("clip_12_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_25_cast_fp16 = real_div(x = x_23_cast_fp16, y = clip_12_cast_fp16)[name = string("xn_25_cast_fp16")];
            fp16 var_23_promoted_12_to_fp16 = const()[name = string("op_23_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1044_cast_fp16 = pow(x = xn_25_cast_fp16, y = var_23_promoted_12_to_fp16)[name = string("op_1044_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_1044_cast_fp16)[name = string("var_25_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1047_cast_fp16 = mul(x = clip_12_cast_fp16, y = clip_12_cast_fp16)[name = string("op_1047_cast_fp16")];
            fp32 var_1048_epsilon_0 = const()[name = string("op_1048_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1048_cast_fp16 = inverse(epsilon = var_1048_epsilon_0, x = var_1047_cast_fp16)[name = string("op_1048_cast_fp16")];
            fp16 var_1049_to_fp16 = const()[name = string("op_1049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1050_cast_fp16 = mul(x = var_1048_cast_fp16, y = var_1049_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1051_cast_fp16 = add(x = var_25_cast_fp16, y = var_1050_cast_fp16)[name = string("op_1051_cast_fp16")];
            fp32 var_1052_epsilon_0 = const()[name = string("op_1052_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1052_cast_fp16 = rsqrt(epsilon = var_1052_epsilon_0, x = var_1051_cast_fp16)[name = string("op_1052_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_25_cast_fp16 = mul(x = xn_25_cast_fp16, y = var_1052_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226103936)))];
            tensor<fp16, [4, 64, 896]> hidden_states_85_cast_fp16 = mul(x = lm_layers_6_input_layernorm_weight_to_fp16, y = out_25_cast_fp16)[name = string("hidden_states_85_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226105792))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226908672))))[name = string("lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226910528)))];
            tensor<fp16, [4, 64, 896]> linear_42_cast_fp16 = linear(bias = lm_layers_6_self_attn_q_proj_bias_to_fp16, weight = lm_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1064 = const()[name = string("op_1064"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1065_cast_fp16 = reshape(shape = var_1064, x = linear_42_cast_fp16)[name = string("op_1065_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = string("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226912384))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027136))))[name = string("lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027456)))];
            tensor<fp16, [4, 64, 128]> linear_43_cast_fp16 = linear(bias = lm_layers_6_self_attn_k_proj_bias_to_fp16, weight = lm_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1070 = const()[name = string("op_1070"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1071_cast_fp16 = reshape(shape = var_1070, x = linear_43_cast_fp16)[name = string("op_1071_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = string("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227027776))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142528))))[name = string("lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227142848)))];
            tensor<fp16, [4, 64, 128]> linear_44_cast_fp16 = linear(bias = lm_layers_6_self_attn_v_proj_bias_to_fp16, weight = lm_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_85_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1076 = const()[name = string("op_1076"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1077_cast_fp16 = reshape(shape = var_1076, x = linear_44_cast_fp16)[name = string("op_1077_cast_fp16")];
            tensor<int32, [4]> hidden_states_91_perm_0 = const()[name = string("hidden_states_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_1065_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [4, 14, 64, 64]> var_1081_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1081_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_13_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_13_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_123_promoted_to_fp16 = const()[name = string("const_123_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1092_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_123_promoted_to_fp16)[name = string("op_1092_cast_fp16")];
            bool var_1094_interleave_0 = const()[name = string("op_1094_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1094_cast_fp16 = concat(axis = var_20, interleave = var_1094_interleave_0, values = (var_1092_cast_fp16, x1_25_cast_fp16))[name = string("op_1094_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1095_cast_fp16 = mul(x = var_1094_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1095_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_13_cast_fp16 = add(x = var_1081_cast_fp16, y = var_1095_cast_fp16)[name = string("query_13_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_1071_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [4, 2, 64, 64]> var_1097_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1097_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_13_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_13_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_126_promoted_to_fp16 = const()[name = string("const_126_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1108_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_126_promoted_to_fp16)[name = string("op_1108_cast_fp16")];
            bool var_1110_interleave_0 = const()[name = string("op_1110_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1110_cast_fp16 = concat(axis = var_20, interleave = var_1110_interleave_0, values = (var_1108_cast_fp16, x1_27_cast_fp16))[name = string("op_1110_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1111_cast_fp16 = mul(x = var_1110_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_87_cast_fp16 = add(x = var_1097_cast_fp16, y = var_1111_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            tensor<int32, [1]> var_1120_axes_0 = const()[name = string("op_1120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1120_cast_fp16 = expand_dims(axes = var_1120_axes_0, x = hidden_states_87_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<int32, [5]> hidden_states_89_reps_0 = const()[name = string("hidden_states_89_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_89_cast_fp16 = tile(reps = hidden_states_89_reps_0, x = var_1120_cast_fp16)[name = string("hidden_states_89_cast_fp16")];
            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_13_cast_fp16 = reshape(shape = var_1128, x = hidden_states_89_cast_fp16)[name = string("key_13_cast_fp16")];
            tensor<int32, [1]> var_1137_axes_0 = const()[name = string("op_1137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_91_cast_fp16 = transpose(perm = hidden_states_91_perm_0, x = var_1077_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1137_cast_fp16 = expand_dims(axes = var_1137_axes_0, x = hidden_states_91_cast_fp16)[name = string("op_1137_cast_fp16")];
            tensor<int32, [5]> hidden_states_93_reps_0 = const()[name = string("hidden_states_93_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_93_cast_fp16 = tile(reps = hidden_states_93_reps_0, x = var_1137_cast_fp16)[name = string("hidden_states_93_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_13_cast_fp16 = reshape(shape = var_1145, x = hidden_states_93_cast_fp16)[name = string("value_13_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_13_cast_fp16 = mul(x = query_13_cast_fp16, y = var_13_to_fp16)[name = string("mul_13_cast_fp16")];
            bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)];
            bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = key_13_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_6_cast_fp16")];
            int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = add_6_cast_fp16)[name = string("softmax_6_cast_fp16")];
            bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)];
            bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")];
            tensor<int32, [4]> var_1148_perm_0 = const()[name = string("op_1148_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1148_cast_fp16 = transpose(perm = var_1148_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [4, 64, 896]> var_1151_cast_fp16 = reshape(shape = var_1150, x = var_1148_cast_fp16)[name = string("op_1151_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227143168))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227946048))))[name = string("lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_45_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = var_1151_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_45_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1157_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<int32, [1]> var_1159_axes_0 = const()[name = string("op_1159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1159_cast_fp16 = reduce_max(axes = var_1159_axes_0, keep_dims = var_12, x = var_1157_cast_fp16)[name = string("op_1159_cast_fp16")];
            fp16 const_135_to_fp16 = const()[name = string("const_135_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_13_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_135_to_fp16, x = var_1159_cast_fp16)[name = string("clip_13_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = clip_13_cast_fp16)[name = string("xn_27_cast_fp16")];
            fp16 var_23_promoted_13_to_fp16 = const()[name = string("op_23_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1162_cast_fp16 = pow(x = xn_27_cast_fp16, y = var_23_promoted_13_to_fp16)[name = string("op_1162_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_1162_cast_fp16)[name = string("var_27_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1165_cast_fp16 = mul(x = clip_13_cast_fp16, y = clip_13_cast_fp16)[name = string("op_1165_cast_fp16")];
            fp32 var_1166_epsilon_0 = const()[name = string("op_1166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1166_cast_fp16 = inverse(epsilon = var_1166_epsilon_0, x = var_1165_cast_fp16)[name = string("op_1166_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1168_cast_fp16 = mul(x = var_1166_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1169_cast_fp16 = add(x = var_27_cast_fp16, y = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            fp32 var_1170_epsilon_0 = const()[name = string("op_1170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1170_cast_fp16 = rsqrt(epsilon = var_1170_epsilon_0, x = var_1169_cast_fp16)[name = string("op_1170_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_27_cast_fp16 = mul(x = xn_27_cast_fp16, y = var_1170_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [896]> lm_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227947904)))];
            tensor<fp16, [4, 64, 896]> input_51_cast_fp16 = mul(x = lm_layers_6_post_attention_layernorm_weight_to_fp16, y = out_27_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227949760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232307968))))[name = string("lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1178_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1178_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232317760))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236675968))))[name = string("lm_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_55_cast_fp16 = mul(x = var_1178_cast_fp16, y = linear_47_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236685760))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241043968))))[name = string("lm_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_48_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_27_cast_fp16 = add(x = x_25_cast_fp16, y = linear_48_cast_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1190_cast_fp16 = abs(x = x_27_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<int32, [1]> var_1192_axes_0 = const()[name = string("op_1192_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1192_cast_fp16 = reduce_max(axes = var_1192_axes_0, keep_dims = var_12, x = var_1190_cast_fp16)[name = string("op_1192_cast_fp16")];
            fp16 const_136_to_fp16 = const()[name = string("const_136_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_14_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_136_to_fp16, x = var_1192_cast_fp16)[name = string("clip_14_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_29_cast_fp16 = real_div(x = x_27_cast_fp16, y = clip_14_cast_fp16)[name = string("xn_29_cast_fp16")];
            fp16 var_23_promoted_14_to_fp16 = const()[name = string("op_23_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1195_cast_fp16 = pow(x = xn_29_cast_fp16, y = var_23_promoted_14_to_fp16)[name = string("op_1195_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_1195_cast_fp16)[name = string("var_29_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1198_cast_fp16 = mul(x = clip_14_cast_fp16, y = clip_14_cast_fp16)[name = string("op_1198_cast_fp16")];
            fp32 var_1199_epsilon_0 = const()[name = string("op_1199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1199_cast_fp16 = inverse(epsilon = var_1199_epsilon_0, x = var_1198_cast_fp16)[name = string("op_1199_cast_fp16")];
            fp16 var_1200_to_fp16 = const()[name = string("op_1200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1201_cast_fp16 = mul(x = var_1199_cast_fp16, y = var_1200_to_fp16)[name = string("op_1201_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1202_cast_fp16 = add(x = var_29_cast_fp16, y = var_1201_cast_fp16)[name = string("op_1202_cast_fp16")];
            fp32 var_1203_epsilon_0 = const()[name = string("op_1203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1203_cast_fp16 = rsqrt(epsilon = var_1203_epsilon_0, x = var_1202_cast_fp16)[name = string("op_1203_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_29_cast_fp16 = mul(x = xn_29_cast_fp16, y = var_1203_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241045824)))];
            tensor<fp16, [4, 64, 896]> hidden_states_99_cast_fp16 = mul(x = lm_layers_7_input_layernorm_weight_to_fp16, y = out_29_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241047680))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241850560))))[name = string("lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241852416)))];
            tensor<fp16, [4, 64, 896]> linear_49_cast_fp16 = linear(bias = lm_layers_7_self_attn_q_proj_bias_to_fp16, weight = lm_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1216_cast_fp16 = reshape(shape = var_1215, x = linear_49_cast_fp16)[name = string("op_1216_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241854272))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969024))))[name = string("lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969344)))];
            tensor<fp16, [4, 64, 128]> linear_50_cast_fp16 = linear(bias = lm_layers_7_self_attn_k_proj_bias_to_fp16, weight = lm_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1222_cast_fp16 = reshape(shape = var_1221, x = linear_50_cast_fp16)[name = string("op_1222_cast_fp16")];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241969664))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084416))))[name = string("lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242084736)))];
            tensor<fp16, [4, 64, 128]> linear_51_cast_fp16 = linear(bias = lm_layers_7_self_attn_v_proj_bias_to_fp16, weight = lm_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_99_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1227 = const()[name = string("op_1227"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1228_cast_fp16 = reshape(shape = var_1227, x = linear_51_cast_fp16)[name = string("op_1228_cast_fp16")];
            tensor<int32, [4]> hidden_states_105_perm_0 = const()[name = string("hidden_states_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_1216_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [4, 14, 64, 64]> var_1232_cast_fp16 = mul(x = q_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1232_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_15_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_15_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_141_promoted_to_fp16 = const()[name = string("const_141_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1243_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_141_promoted_to_fp16)[name = string("op_1243_cast_fp16")];
            bool var_1245_interleave_0 = const()[name = string("op_1245_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1245_cast_fp16 = concat(axis = var_20, interleave = var_1245_interleave_0, values = (var_1243_cast_fp16, x1_29_cast_fp16))[name = string("op_1245_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1246_cast_fp16 = mul(x = var_1245_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1246_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_15_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1246_cast_fp16)[name = string("query_15_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_1222_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [4, 2, 64, 64]> var_1248_cast_fp16 = mul(x = k_15_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1248_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_15_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_15_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_144_promoted_to_fp16 = const()[name = string("const_144_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1259_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_144_promoted_to_fp16)[name = string("op_1259_cast_fp16")];
            bool var_1261_interleave_0 = const()[name = string("op_1261_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1261_cast_fp16 = concat(axis = var_20, interleave = var_1261_interleave_0, values = (var_1259_cast_fp16, x1_31_cast_fp16))[name = string("op_1261_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1262_cast_fp16 = mul(x = var_1261_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1262_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_101_cast_fp16 = add(x = var_1248_cast_fp16, y = var_1262_cast_fp16)[name = string("hidden_states_101_cast_fp16")];
            tensor<int32, [1]> var_1271_axes_0 = const()[name = string("op_1271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1271_cast_fp16 = expand_dims(axes = var_1271_axes_0, x = hidden_states_101_cast_fp16)[name = string("op_1271_cast_fp16")];
            tensor<int32, [5]> hidden_states_103_reps_0 = const()[name = string("hidden_states_103_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_103_cast_fp16 = tile(reps = hidden_states_103_reps_0, x = var_1271_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            tensor<int32, [4]> var_1279 = const()[name = string("op_1279"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_15_cast_fp16 = reshape(shape = var_1279, x = hidden_states_103_cast_fp16)[name = string("key_15_cast_fp16")];
            tensor<int32, [1]> var_1288_axes_0 = const()[name = string("op_1288_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_105_cast_fp16 = transpose(perm = hidden_states_105_perm_0, x = var_1228_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1288_cast_fp16 = expand_dims(axes = var_1288_axes_0, x = hidden_states_105_cast_fp16)[name = string("op_1288_cast_fp16")];
            tensor<int32, [5]> hidden_states_107_reps_0 = const()[name = string("hidden_states_107_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_107_cast_fp16 = tile(reps = hidden_states_107_reps_0, x = var_1288_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<int32, [4]> var_1296 = const()[name = string("op_1296"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_15_cast_fp16 = reshape(shape = var_1296, x = hidden_states_107_cast_fp16)[name = string("value_15_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_15_cast_fp16 = mul(x = query_15_cast_fp16, y = var_13_to_fp16)[name = string("mul_15_cast_fp16")];
            bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)];
            bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = key_15_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_7_cast_fp16")];
            int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = add_7_cast_fp16)[name = string("softmax_7_cast_fp16")];
            bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)];
            bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")];
            tensor<int32, [4]> var_1299_perm_0 = const()[name = string("op_1299_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1299_cast_fp16 = transpose(perm = var_1299_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [4, 64, 896]> var_1302_cast_fp16 = reshape(shape = var_1301, x = var_1299_cast_fp16)[name = string("op_1302_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242085056))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242887936))))[name = string("lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_52_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = var_1302_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_29_cast_fp16 = add(x = x_27_cast_fp16, y = linear_52_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1308_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1308_cast_fp16")];
            tensor<int32, [1]> var_1310_axes_0 = const()[name = string("op_1310_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1310_cast_fp16 = reduce_max(axes = var_1310_axes_0, keep_dims = var_12, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 const_153_to_fp16 = const()[name = string("const_153_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_15_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_153_to_fp16, x = var_1310_cast_fp16)[name = string("clip_15_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = clip_15_cast_fp16)[name = string("xn_31_cast_fp16")];
            fp16 var_23_promoted_15_to_fp16 = const()[name = string("op_23_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1313_cast_fp16 = pow(x = xn_31_cast_fp16, y = var_23_promoted_15_to_fp16)[name = string("op_1313_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_1313_cast_fp16)[name = string("var_31_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1316_cast_fp16 = mul(x = clip_15_cast_fp16, y = clip_15_cast_fp16)[name = string("op_1316_cast_fp16")];
            fp32 var_1317_epsilon_0 = const()[name = string("op_1317_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1317_cast_fp16 = inverse(epsilon = var_1317_epsilon_0, x = var_1316_cast_fp16)[name = string("op_1317_cast_fp16")];
            fp16 var_1318_to_fp16 = const()[name = string("op_1318_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1319_cast_fp16 = mul(x = var_1317_cast_fp16, y = var_1318_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1320_cast_fp16 = add(x = var_31_cast_fp16, y = var_1319_cast_fp16)[name = string("op_1320_cast_fp16")];
            fp32 var_1321_epsilon_0 = const()[name = string("op_1321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1321_cast_fp16 = rsqrt(epsilon = var_1321_epsilon_0, x = var_1320_cast_fp16)[name = string("op_1321_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_31_cast_fp16 = mul(x = xn_31_cast_fp16, y = var_1321_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [896]> lm_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242889792)))];
            tensor<fp16, [4, 64, 896]> input_59_cast_fp16 = mul(x = lm_layers_7_post_attention_layernorm_weight_to_fp16, y = out_31_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(242891648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247249856))))[name = string("lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247259648))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251617856))))[name = string("lm_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_63_cast_fp16 = mul(x = var_1329_cast_fp16, y = linear_54_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251627648))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255985856))))[name = string("lm_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_55_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1341_cast_fp16 = abs(x = x_31_cast_fp16)[name = string("op_1341_cast_fp16")];
            tensor<int32, [1]> var_1343_axes_0 = const()[name = string("op_1343_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1343_cast_fp16 = reduce_max(axes = var_1343_axes_0, keep_dims = var_12, x = var_1341_cast_fp16)[name = string("op_1343_cast_fp16")];
            fp16 const_154_to_fp16 = const()[name = string("const_154_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_16_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_154_to_fp16, x = var_1343_cast_fp16)[name = string("clip_16_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_33_cast_fp16 = real_div(x = x_31_cast_fp16, y = clip_16_cast_fp16)[name = string("xn_33_cast_fp16")];
            fp16 var_23_promoted_16_to_fp16 = const()[name = string("op_23_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1346_cast_fp16 = pow(x = xn_33_cast_fp16, y = var_23_promoted_16_to_fp16)[name = string("op_1346_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_1346_cast_fp16)[name = string("var_33_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1349_cast_fp16 = mul(x = clip_16_cast_fp16, y = clip_16_cast_fp16)[name = string("op_1349_cast_fp16")];
            fp32 var_1350_epsilon_0 = const()[name = string("op_1350_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1350_cast_fp16 = inverse(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = string("op_1350_cast_fp16")];
            fp16 var_1351_to_fp16 = const()[name = string("op_1351_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1352_cast_fp16 = mul(x = var_1350_cast_fp16, y = var_1351_to_fp16)[name = string("op_1352_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1353_cast_fp16 = add(x = var_33_cast_fp16, y = var_1352_cast_fp16)[name = string("op_1353_cast_fp16")];
            fp32 var_1354_epsilon_0 = const()[name = string("op_1354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1354_cast_fp16 = rsqrt(epsilon = var_1354_epsilon_0, x = var_1353_cast_fp16)[name = string("op_1354_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_33_cast_fp16 = mul(x = xn_33_cast_fp16, y = var_1354_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255987712)))];
            tensor<fp16, [4, 64, 896]> hidden_states_113_cast_fp16 = mul(x = lm_layers_8_input_layernorm_weight_to_fp16, y = out_33_cast_fp16)[name = string("hidden_states_113_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255989568))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256792448))))[name = string("lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256794304)))];
            tensor<fp16, [4, 64, 896]> linear_56_cast_fp16 = linear(bias = lm_layers_8_self_attn_q_proj_bias_to_fp16, weight = lm_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1366 = const()[name = string("op_1366"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1367_cast_fp16 = reshape(shape = var_1366, x = linear_56_cast_fp16)[name = string("op_1367_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = string("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256796160))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256910912))))[name = string("lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911232)))];
            tensor<fp16, [4, 64, 128]> linear_57_cast_fp16 = linear(bias = lm_layers_8_self_attn_k_proj_bias_to_fp16, weight = lm_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1372 = const()[name = string("op_1372"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1373_cast_fp16 = reshape(shape = var_1372, x = linear_57_cast_fp16)[name = string("op_1373_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256911552))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026304))))[name = string("lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026624)))];
            tensor<fp16, [4, 64, 128]> linear_58_cast_fp16 = linear(bias = lm_layers_8_self_attn_v_proj_bias_to_fp16, weight = lm_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_113_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1378 = const()[name = string("op_1378"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1379_cast_fp16 = reshape(shape = var_1378, x = linear_58_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<int32, [4]> hidden_states_119_perm_0 = const()[name = string("hidden_states_119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_1367_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [4, 14, 64, 64]> var_1383_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1383_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_17_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_17_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_159_promoted_to_fp16 = const()[name = string("const_159_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1394_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_159_promoted_to_fp16)[name = string("op_1394_cast_fp16")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1396_cast_fp16 = concat(axis = var_20, interleave = var_1396_interleave_0, values = (var_1394_cast_fp16, x1_33_cast_fp16))[name = string("op_1396_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1397_cast_fp16 = mul(x = var_1396_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1397_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_17_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1397_cast_fp16)[name = string("query_17_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_1373_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [4, 2, 64, 64]> var_1399_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1399_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_17_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_17_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_162_promoted_to_fp16 = const()[name = string("const_162_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1410_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_162_promoted_to_fp16)[name = string("op_1410_cast_fp16")];
            bool var_1412_interleave_0 = const()[name = string("op_1412_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1412_cast_fp16 = concat(axis = var_20, interleave = var_1412_interleave_0, values = (var_1410_cast_fp16, x1_35_cast_fp16))[name = string("op_1412_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1413_cast_fp16 = mul(x = var_1412_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_115_cast_fp16 = add(x = var_1399_cast_fp16, y = var_1413_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<int32, [1]> var_1422_axes_0 = const()[name = string("op_1422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1422_cast_fp16 = expand_dims(axes = var_1422_axes_0, x = hidden_states_115_cast_fp16)[name = string("op_1422_cast_fp16")];
            tensor<int32, [5]> hidden_states_117_reps_0 = const()[name = string("hidden_states_117_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_117_cast_fp16 = tile(reps = hidden_states_117_reps_0, x = var_1422_cast_fp16)[name = string("hidden_states_117_cast_fp16")];
            tensor<int32, [4]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_17_cast_fp16 = reshape(shape = var_1430, x = hidden_states_117_cast_fp16)[name = string("key_17_cast_fp16")];
            tensor<int32, [1]> var_1439_axes_0 = const()[name = string("op_1439_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_119_cast_fp16 = transpose(perm = hidden_states_119_perm_0, x = var_1379_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1439_cast_fp16 = expand_dims(axes = var_1439_axes_0, x = hidden_states_119_cast_fp16)[name = string("op_1439_cast_fp16")];
            tensor<int32, [5]> hidden_states_121_reps_0 = const()[name = string("hidden_states_121_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_121_cast_fp16 = tile(reps = hidden_states_121_reps_0, x = var_1439_cast_fp16)[name = string("hidden_states_121_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_17_cast_fp16 = reshape(shape = var_1447, x = hidden_states_121_cast_fp16)[name = string("value_17_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_17_cast_fp16 = mul(x = query_17_cast_fp16, y = var_13_to_fp16)[name = string("mul_17_cast_fp16")];
            bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)];
            bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = key_17_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_8_cast_fp16")];
            int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = add_8_cast_fp16)[name = string("softmax_8_cast_fp16")];
            bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)];
            bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")];
            tensor<int32, [4]> var_1450_perm_0 = const()[name = string("op_1450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1450_cast_fp16 = transpose(perm = var_1450_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [4, 64, 896]> var_1453_cast_fp16 = reshape(shape = var_1452, x = var_1450_cast_fp16)[name = string("op_1453_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257026944))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257829824))))[name = string("lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = var_1453_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_33_cast_fp16 = add(x = x_31_cast_fp16, y = linear_59_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1459_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_1459_cast_fp16")];
            tensor<int32, [1]> var_1461_axes_0 = const()[name = string("op_1461_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1461_cast_fp16 = reduce_max(axes = var_1461_axes_0, keep_dims = var_12, x = var_1459_cast_fp16)[name = string("op_1461_cast_fp16")];
            fp16 const_171_to_fp16 = const()[name = string("const_171_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_17_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_171_to_fp16, x = var_1461_cast_fp16)[name = string("clip_17_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_35_cast_fp16 = real_div(x = x_33_cast_fp16, y = clip_17_cast_fp16)[name = string("xn_35_cast_fp16")];
            fp16 var_23_promoted_17_to_fp16 = const()[name = string("op_23_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1464_cast_fp16 = pow(x = xn_35_cast_fp16, y = var_23_promoted_17_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_1464_cast_fp16)[name = string("var_35_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1467_cast_fp16 = mul(x = clip_17_cast_fp16, y = clip_17_cast_fp16)[name = string("op_1467_cast_fp16")];
            fp32 var_1468_epsilon_0 = const()[name = string("op_1468_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1468_cast_fp16 = inverse(epsilon = var_1468_epsilon_0, x = var_1467_cast_fp16)[name = string("op_1468_cast_fp16")];
            fp16 var_1469_to_fp16 = const()[name = string("op_1469_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1470_cast_fp16 = mul(x = var_1468_cast_fp16, y = var_1469_to_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1471_cast_fp16 = add(x = var_35_cast_fp16, y = var_1470_cast_fp16)[name = string("op_1471_cast_fp16")];
            fp32 var_1472_epsilon_0 = const()[name = string("op_1472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1472_cast_fp16 = rsqrt(epsilon = var_1472_epsilon_0, x = var_1471_cast_fp16)[name = string("op_1472_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_35_cast_fp16 = mul(x = xn_35_cast_fp16, y = var_1472_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [896]> lm_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257831680)))];
            tensor<fp16, [4, 64, 896]> input_67_cast_fp16 = mul(x = lm_layers_8_post_attention_layernorm_weight_to_fp16, y = out_35_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257833536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262191744))))[name = string("lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1480_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1480_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262201536))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266559744))))[name = string("lm_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_71_cast_fp16 = mul(x = var_1480_cast_fp16, y = linear_61_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266569536))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270927744))))[name = string("lm_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_62_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_35_cast_fp16 = add(x = x_33_cast_fp16, y = linear_62_cast_fp16)[name = string("x_35_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1492_cast_fp16 = abs(x = x_35_cast_fp16)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> var_1494_axes_0 = const()[name = string("op_1494_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1494_cast_fp16 = reduce_max(axes = var_1494_axes_0, keep_dims = var_12, x = var_1492_cast_fp16)[name = string("op_1494_cast_fp16")];
            fp16 const_172_to_fp16 = const()[name = string("const_172_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_18_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_172_to_fp16, x = var_1494_cast_fp16)[name = string("clip_18_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_37_cast_fp16 = real_div(x = x_35_cast_fp16, y = clip_18_cast_fp16)[name = string("xn_37_cast_fp16")];
            fp16 var_23_promoted_18_to_fp16 = const()[name = string("op_23_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1497_cast_fp16 = pow(x = xn_37_cast_fp16, y = var_23_promoted_18_to_fp16)[name = string("op_1497_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_1497_cast_fp16)[name = string("var_37_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1500_cast_fp16 = mul(x = clip_18_cast_fp16, y = clip_18_cast_fp16)[name = string("op_1500_cast_fp16")];
            fp32 var_1501_epsilon_0 = const()[name = string("op_1501_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1501_cast_fp16 = inverse(epsilon = var_1501_epsilon_0, x = var_1500_cast_fp16)[name = string("op_1501_cast_fp16")];
            fp16 var_1502_to_fp16 = const()[name = string("op_1502_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1503_cast_fp16 = mul(x = var_1501_cast_fp16, y = var_1502_to_fp16)[name = string("op_1503_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1504_cast_fp16 = add(x = var_37_cast_fp16, y = var_1503_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp32 var_1505_epsilon_0 = const()[name = string("op_1505_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1505_cast_fp16 = rsqrt(epsilon = var_1505_epsilon_0, x = var_1504_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_37_cast_fp16 = mul(x = xn_37_cast_fp16, y = var_1505_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270929600)))];
            tensor<fp16, [4, 64, 896]> hidden_states_127_cast_fp16 = mul(x = lm_layers_9_input_layernorm_weight_to_fp16, y = out_37_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270931456))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271734336))))[name = string("lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271736192)))];
            tensor<fp16, [4, 64, 896]> linear_63_cast_fp16 = linear(bias = lm_layers_9_self_attn_q_proj_bias_to_fp16, weight = lm_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1518_cast_fp16 = reshape(shape = var_1517, x = linear_63_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271738048))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271852800))))[name = string("lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853120)))];
            tensor<fp16, [4, 64, 128]> linear_64_cast_fp16 = linear(bias = lm_layers_9_self_attn_k_proj_bias_to_fp16, weight = lm_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1523 = const()[name = string("op_1523"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1524_cast_fp16 = reshape(shape = var_1523, x = linear_64_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271853440))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968192))))[name = string("lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968512)))];
            tensor<fp16, [4, 64, 128]> linear_65_cast_fp16 = linear(bias = lm_layers_9_self_attn_v_proj_bias_to_fp16, weight = lm_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_127_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1530_cast_fp16 = reshape(shape = var_1529, x = linear_65_cast_fp16)[name = string("op_1530_cast_fp16")];
            tensor<int32, [4]> hidden_states_133_perm_0 = const()[name = string("hidden_states_133_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_19_cast_fp16 = transpose(perm = q_19_perm_0, x = var_1518_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [4, 14, 64, 64]> var_1534_cast_fp16 = mul(x = q_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_19_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_19_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_177_promoted_to_fp16 = const()[name = string("const_177_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1545_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_177_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            bool var_1547_interleave_0 = const()[name = string("op_1547_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1547_cast_fp16 = concat(axis = var_20, interleave = var_1547_interleave_0, values = (var_1545_cast_fp16, x1_37_cast_fp16))[name = string("op_1547_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1548_cast_fp16 = mul(x = var_1547_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1548_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_19_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1548_cast_fp16)[name = string("query_19_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = var_1524_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [4, 2, 64, 64]> var_1550_cast_fp16 = mul(x = k_19_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_19_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_19_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_180_promoted_to_fp16 = const()[name = string("const_180_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1561_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_180_promoted_to_fp16)[name = string("op_1561_cast_fp16")];
            bool var_1563_interleave_0 = const()[name = string("op_1563_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1563_cast_fp16 = concat(axis = var_20, interleave = var_1563_interleave_0, values = (var_1561_cast_fp16, x1_39_cast_fp16))[name = string("op_1563_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1564_cast_fp16 = mul(x = var_1563_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_129_cast_fp16 = add(x = var_1550_cast_fp16, y = var_1564_cast_fp16)[name = string("hidden_states_129_cast_fp16")];
            tensor<int32, [1]> var_1573_axes_0 = const()[name = string("op_1573_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1573_cast_fp16 = expand_dims(axes = var_1573_axes_0, x = hidden_states_129_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<int32, [5]> hidden_states_131_reps_0 = const()[name = string("hidden_states_131_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_131_cast_fp16 = tile(reps = hidden_states_131_reps_0, x = var_1573_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<int32, [4]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_19_cast_fp16 = reshape(shape = var_1581, x = hidden_states_131_cast_fp16)[name = string("key_19_cast_fp16")];
            tensor<int32, [1]> var_1590_axes_0 = const()[name = string("op_1590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_133_cast_fp16 = transpose(perm = hidden_states_133_perm_0, x = var_1530_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1590_cast_fp16 = expand_dims(axes = var_1590_axes_0, x = hidden_states_133_cast_fp16)[name = string("op_1590_cast_fp16")];
            tensor<int32, [5]> hidden_states_135_reps_0 = const()[name = string("hidden_states_135_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_135_cast_fp16 = tile(reps = hidden_states_135_reps_0, x = var_1590_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            tensor<int32, [4]> var_1598 = const()[name = string("op_1598"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_19_cast_fp16 = reshape(shape = var_1598, x = hidden_states_135_cast_fp16)[name = string("value_19_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_19_cast_fp16 = mul(x = query_19_cast_fp16, y = var_13_to_fp16)[name = string("mul_19_cast_fp16")];
            bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)];
            bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = key_19_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_9_cast_fp16")];
            int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = add_9_cast_fp16)[name = string("softmax_9_cast_fp16")];
            bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)];
            bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")];
            tensor<int32, [4]> var_1601_perm_0 = const()[name = string("op_1601_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1601_cast_fp16 = transpose(perm = var_1601_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [4, 64, 896]> var_1604_cast_fp16 = reshape(shape = var_1603, x = var_1601_cast_fp16)[name = string("op_1604_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271968832))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272771712))))[name = string("lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_66_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = var_1604_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_66_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1610_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_1610_cast_fp16")];
            tensor<int32, [1]> var_1612_axes_0 = const()[name = string("op_1612_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1612_cast_fp16 = reduce_max(axes = var_1612_axes_0, keep_dims = var_12, x = var_1610_cast_fp16)[name = string("op_1612_cast_fp16")];
            fp16 const_189_to_fp16 = const()[name = string("const_189_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_19_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_189_to_fp16, x = var_1612_cast_fp16)[name = string("clip_19_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = clip_19_cast_fp16)[name = string("xn_39_cast_fp16")];
            fp16 var_23_promoted_19_to_fp16 = const()[name = string("op_23_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1615_cast_fp16 = pow(x = xn_39_cast_fp16, y = var_23_promoted_19_to_fp16)[name = string("op_1615_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1615_cast_fp16)[name = string("var_39_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1618_cast_fp16 = mul(x = clip_19_cast_fp16, y = clip_19_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1619_cast_fp16 = inverse(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1622_cast_fp16 = add(x = var_39_cast_fp16, y = var_1621_cast_fp16)[name = string("op_1622_cast_fp16")];
            fp32 var_1623_epsilon_0 = const()[name = string("op_1623_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1623_cast_fp16 = rsqrt(epsilon = var_1623_epsilon_0, x = var_1622_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_39_cast_fp16 = mul(x = xn_39_cast_fp16, y = var_1623_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [896]> lm_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272773568)))];
            tensor<fp16, [4, 64, 896]> input_75_cast_fp16 = mul(x = lm_layers_9_post_attention_layernorm_weight_to_fp16, y = out_39_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272775424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277133632))))[name = string("lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1631_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_1631_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277143424))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281501632))))[name = string("lm_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_79_cast_fp16 = mul(x = var_1631_cast_fp16, y = linear_68_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281511424))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285869632))))[name = string("lm_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_69_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_39_cast_fp16 = add(x = x_37_cast_fp16, y = linear_69_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1643_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_1643_cast_fp16")];
            tensor<int32, [1]> var_1645_axes_0 = const()[name = string("op_1645_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1645_cast_fp16 = reduce_max(axes = var_1645_axes_0, keep_dims = var_12, x = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp16 const_190_to_fp16 = const()[name = string("const_190_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_20_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_190_to_fp16, x = var_1645_cast_fp16)[name = string("clip_20_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_41_cast_fp16 = real_div(x = x_39_cast_fp16, y = clip_20_cast_fp16)[name = string("xn_41_cast_fp16")];
            fp16 var_23_promoted_20_to_fp16 = const()[name = string("op_23_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1648_cast_fp16 = pow(x = xn_41_cast_fp16, y = var_23_promoted_20_to_fp16)[name = string("op_1648_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1648_cast_fp16)[name = string("var_41_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1651_cast_fp16 = mul(x = clip_20_cast_fp16, y = clip_20_cast_fp16)[name = string("op_1651_cast_fp16")];
            fp32 var_1652_epsilon_0 = const()[name = string("op_1652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1652_cast_fp16 = inverse(epsilon = var_1652_epsilon_0, x = var_1651_cast_fp16)[name = string("op_1652_cast_fp16")];
            fp16 var_1653_to_fp16 = const()[name = string("op_1653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1654_cast_fp16 = mul(x = var_1652_cast_fp16, y = var_1653_to_fp16)[name = string("op_1654_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1655_cast_fp16 = add(x = var_41_cast_fp16, y = var_1654_cast_fp16)[name = string("op_1655_cast_fp16")];
            fp32 var_1656_epsilon_0 = const()[name = string("op_1656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1656_cast_fp16 = rsqrt(epsilon = var_1656_epsilon_0, x = var_1655_cast_fp16)[name = string("op_1656_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_41_cast_fp16 = mul(x = xn_41_cast_fp16, y = var_1656_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285871488)))];
            tensor<fp16, [4, 64, 896]> hidden_states_141_cast_fp16 = mul(x = lm_layers_10_input_layernorm_weight_to_fp16, y = out_41_cast_fp16)[name = string("hidden_states_141_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285873344))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286676224))))[name = string("lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286678080)))];
            tensor<fp16, [4, 64, 896]> linear_70_cast_fp16 = linear(bias = lm_layers_10_self_attn_q_proj_bias_to_fp16, weight = lm_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1669_cast_fp16 = reshape(shape = var_1668, x = linear_70_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286679936))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286794688))))[name = string("lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795008)))];
            tensor<fp16, [4, 64, 128]> linear_71_cast_fp16 = linear(bias = lm_layers_10_self_attn_k_proj_bias_to_fp16, weight = lm_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1674 = const()[name = string("op_1674"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1675_cast_fp16 = reshape(shape = var_1674, x = linear_71_cast_fp16)[name = string("op_1675_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = string("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286795328))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910080))))[name = string("lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910400)))];
            tensor<fp16, [4, 64, 128]> linear_72_cast_fp16 = linear(bias = lm_layers_10_self_attn_v_proj_bias_to_fp16, weight = lm_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_141_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1680 = const()[name = string("op_1680"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1681_cast_fp16 = reshape(shape = var_1680, x = linear_72_cast_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> hidden_states_147_perm_0 = const()[name = string("hidden_states_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1669_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [4, 14, 64, 64]> var_1685_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_21_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_21_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_195_promoted_to_fp16 = const()[name = string("const_195_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1696_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_195_promoted_to_fp16)[name = string("op_1696_cast_fp16")];
            bool var_1698_interleave_0 = const()[name = string("op_1698_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1698_cast_fp16 = concat(axis = var_20, interleave = var_1698_interleave_0, values = (var_1696_cast_fp16, x1_41_cast_fp16))[name = string("op_1698_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1699_cast_fp16 = mul(x = var_1698_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_21_cast_fp16 = add(x = var_1685_cast_fp16, y = var_1699_cast_fp16)[name = string("query_21_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1675_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [4, 2, 64, 64]> var_1701_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_21_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_21_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_198_promoted_to_fp16 = const()[name = string("const_198_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1712_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_198_promoted_to_fp16)[name = string("op_1712_cast_fp16")];
            bool var_1714_interleave_0 = const()[name = string("op_1714_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1714_cast_fp16 = concat(axis = var_20, interleave = var_1714_interleave_0, values = (var_1712_cast_fp16, x1_43_cast_fp16))[name = string("op_1714_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1715_cast_fp16 = mul(x = var_1714_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_143_cast_fp16 = add(x = var_1701_cast_fp16, y = var_1715_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [1]> var_1724_axes_0 = const()[name = string("op_1724_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1724_cast_fp16 = expand_dims(axes = var_1724_axes_0, x = hidden_states_143_cast_fp16)[name = string("op_1724_cast_fp16")];
            tensor<int32, [5]> hidden_states_145_reps_0 = const()[name = string("hidden_states_145_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_145_cast_fp16 = tile(reps = hidden_states_145_reps_0, x = var_1724_cast_fp16)[name = string("hidden_states_145_cast_fp16")];
            tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_21_cast_fp16 = reshape(shape = var_1732, x = hidden_states_145_cast_fp16)[name = string("key_21_cast_fp16")];
            tensor<int32, [1]> var_1741_axes_0 = const()[name = string("op_1741_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_147_cast_fp16 = transpose(perm = hidden_states_147_perm_0, x = var_1681_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1741_cast_fp16 = expand_dims(axes = var_1741_axes_0, x = hidden_states_147_cast_fp16)[name = string("op_1741_cast_fp16")];
            tensor<int32, [5]> hidden_states_149_reps_0 = const()[name = string("hidden_states_149_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_149_cast_fp16 = tile(reps = hidden_states_149_reps_0, x = var_1741_cast_fp16)[name = string("hidden_states_149_cast_fp16")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_21_cast_fp16 = reshape(shape = var_1749, x = hidden_states_149_cast_fp16)[name = string("value_21_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_21_cast_fp16 = mul(x = query_21_cast_fp16, y = var_13_to_fp16)[name = string("mul_21_cast_fp16")];
            bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)];
            bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = key_21_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_10_cast_fp16")];
            int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = add_10_cast_fp16)[name = string("softmax_10_cast_fp16")];
            bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)];
            bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")];
            tensor<int32, [4]> var_1752_perm_0 = const()[name = string("op_1752_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1754 = const()[name = string("op_1754"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1752_cast_fp16 = transpose(perm = var_1752_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [4, 64, 896]> var_1755_cast_fp16 = reshape(shape = var_1754, x = var_1752_cast_fp16)[name = string("op_1755_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286910720))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287713600))))[name = string("lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_73_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = var_1755_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_41_cast_fp16 = add(x = x_39_cast_fp16, y = linear_73_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1761_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_1761_cast_fp16")];
            tensor<int32, [1]> var_1763_axes_0 = const()[name = string("op_1763_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1763_cast_fp16 = reduce_max(axes = var_1763_axes_0, keep_dims = var_12, x = var_1761_cast_fp16)[name = string("op_1763_cast_fp16")];
            fp16 const_207_to_fp16 = const()[name = string("const_207_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_21_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_207_to_fp16, x = var_1763_cast_fp16)[name = string("clip_21_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = clip_21_cast_fp16)[name = string("xn_43_cast_fp16")];
            fp16 var_23_promoted_21_to_fp16 = const()[name = string("op_23_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1766_cast_fp16 = pow(x = xn_43_cast_fp16, y = var_23_promoted_21_to_fp16)[name = string("op_1766_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1766_cast_fp16)[name = string("var_43_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1769_cast_fp16 = mul(x = clip_21_cast_fp16, y = clip_21_cast_fp16)[name = string("op_1769_cast_fp16")];
            fp32 var_1770_epsilon_0 = const()[name = string("op_1770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1770_cast_fp16 = inverse(epsilon = var_1770_epsilon_0, x = var_1769_cast_fp16)[name = string("op_1770_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1772_cast_fp16 = mul(x = var_1770_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1773_cast_fp16 = add(x = var_43_cast_fp16, y = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            fp32 var_1774_epsilon_0 = const()[name = string("op_1774_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1774_cast_fp16 = rsqrt(epsilon = var_1774_epsilon_0, x = var_1773_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_43_cast_fp16 = mul(x = xn_43_cast_fp16, y = var_1774_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [896]> lm_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287715456)))];
            tensor<fp16, [4, 64, 896]> input_83_cast_fp16 = mul(x = lm_layers_10_post_attention_layernorm_weight_to_fp16, y = out_43_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287717312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292075520))))[name = string("lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1782_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_1782_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292085312))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296443520))))[name = string("lm_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_87_cast_fp16 = mul(x = var_1782_cast_fp16, y = linear_75_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296453312))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300811520))))[name = string("lm_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_76_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_76_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1794_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_1794_cast_fp16")];
            tensor<int32, [1]> var_1796_axes_0 = const()[name = string("op_1796_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1796_cast_fp16 = reduce_max(axes = var_1796_axes_0, keep_dims = var_12, x = var_1794_cast_fp16)[name = string("op_1796_cast_fp16")];
            fp16 const_208_to_fp16 = const()[name = string("const_208_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_22_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_208_to_fp16, x = var_1796_cast_fp16)[name = string("clip_22_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_45_cast_fp16 = real_div(x = x_43_cast_fp16, y = clip_22_cast_fp16)[name = string("xn_45_cast_fp16")];
            fp16 var_23_promoted_22_to_fp16 = const()[name = string("op_23_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1799_cast_fp16 = pow(x = xn_45_cast_fp16, y = var_23_promoted_22_to_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1799_cast_fp16)[name = string("var_45_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1802_cast_fp16 = mul(x = clip_22_cast_fp16, y = clip_22_cast_fp16)[name = string("op_1802_cast_fp16")];
            fp32 var_1803_epsilon_0 = const()[name = string("op_1803_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1803_cast_fp16 = inverse(epsilon = var_1803_epsilon_0, x = var_1802_cast_fp16)[name = string("op_1803_cast_fp16")];
            fp16 var_1804_to_fp16 = const()[name = string("op_1804_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1805_cast_fp16 = mul(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1806_cast_fp16 = add(x = var_45_cast_fp16, y = var_1805_cast_fp16)[name = string("op_1806_cast_fp16")];
            fp32 var_1807_epsilon_0 = const()[name = string("op_1807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1807_cast_fp16 = rsqrt(epsilon = var_1807_epsilon_0, x = var_1806_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_45_cast_fp16 = mul(x = xn_45_cast_fp16, y = var_1807_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300813376)))];
            tensor<fp16, [4, 64, 896]> hidden_states_155_cast_fp16 = mul(x = lm_layers_11_input_layernorm_weight_to_fp16, y = out_45_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300815232))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301618112))))[name = string("lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301619968)))];
            tensor<fp16, [4, 64, 896]> linear_77_cast_fp16 = linear(bias = lm_layers_11_self_attn_q_proj_bias_to_fp16, weight = lm_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_1819 = const()[name = string("op_1819"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1820_cast_fp16 = reshape(shape = var_1819, x = linear_77_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<int32, [4]> q_23_perm_0 = const()[name = string("q_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301621824))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736576))))[name = string("lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301736896)))];
            tensor<fp16, [4, 64, 128]> linear_78_cast_fp16 = linear(bias = lm_layers_11_self_attn_k_proj_bias_to_fp16, weight = lm_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1826_cast_fp16 = reshape(shape = var_1825, x = linear_78_cast_fp16)[name = string("op_1826_cast_fp16")];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301737216))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301851968))))[name = string("lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852288)))];
            tensor<fp16, [4, 64, 128]> linear_79_cast_fp16 = linear(bias = lm_layers_11_self_attn_v_proj_bias_to_fp16, weight = lm_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_155_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_1831 = const()[name = string("op_1831"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1832_cast_fp16 = reshape(shape = var_1831, x = linear_79_cast_fp16)[name = string("op_1832_cast_fp16")];
            tensor<int32, [4]> hidden_states_161_perm_0 = const()[name = string("hidden_states_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_23_cast_fp16 = transpose(perm = q_23_perm_0, x = var_1820_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [4, 14, 64, 64]> var_1836_cast_fp16 = mul(x = q_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_23_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_23_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_213_promoted_to_fp16 = const()[name = string("const_213_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1847_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_213_promoted_to_fp16)[name = string("op_1847_cast_fp16")];
            bool var_1849_interleave_0 = const()[name = string("op_1849_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_1849_cast_fp16 = concat(axis = var_20, interleave = var_1849_interleave_0, values = (var_1847_cast_fp16, x1_45_cast_fp16))[name = string("op_1849_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_1850_cast_fp16 = mul(x = var_1849_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1850_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_23_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1850_cast_fp16)[name = string("query_23_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = var_1826_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [4, 2, 64, 64]> var_1852_cast_fp16 = mul(x = k_23_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_23_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_23_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_216_promoted_to_fp16 = const()[name = string("const_216_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_1863_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_216_promoted_to_fp16)[name = string("op_1863_cast_fp16")];
            bool var_1865_interleave_0 = const()[name = string("op_1865_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_1865_cast_fp16 = concat(axis = var_20, interleave = var_1865_interleave_0, values = (var_1863_cast_fp16, x1_47_cast_fp16))[name = string("op_1865_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_1866_cast_fp16 = mul(x = var_1865_cast_fp16, y = sin_5_cast_fp16)[name = string("op_1866_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_157_cast_fp16 = add(x = var_1852_cast_fp16, y = var_1866_cast_fp16)[name = string("hidden_states_157_cast_fp16")];
            tensor<int32, [1]> var_1875_axes_0 = const()[name = string("op_1875_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1875_cast_fp16 = expand_dims(axes = var_1875_axes_0, x = hidden_states_157_cast_fp16)[name = string("op_1875_cast_fp16")];
            tensor<int32, [5]> hidden_states_159_reps_0 = const()[name = string("hidden_states_159_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_159_cast_fp16 = tile(reps = hidden_states_159_reps_0, x = var_1875_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_23_cast_fp16 = reshape(shape = var_1883, x = hidden_states_159_cast_fp16)[name = string("key_23_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_161_cast_fp16 = transpose(perm = hidden_states_161_perm_0, x = var_1832_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_1892_cast_fp16 = expand_dims(axes = var_1892_axes_0, x = hidden_states_161_cast_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [5]> hidden_states_163_reps_0 = const()[name = string("hidden_states_163_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_163_cast_fp16 = tile(reps = hidden_states_163_reps_0, x = var_1892_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<int32, [4]> var_1900 = const()[name = string("op_1900"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_23_cast_fp16 = reshape(shape = var_1900, x = hidden_states_163_cast_fp16)[name = string("value_23_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_23_cast_fp16 = mul(x = query_23_cast_fp16, y = var_13_to_fp16)[name = string("mul_23_cast_fp16")];
            bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)];
            bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = key_23_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_11_cast_fp16")];
            int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = add_11_cast_fp16)[name = string("softmax_11_cast_fp16")];
            bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)];
            bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_23_cast_fp16)[name = string("attn_output_45_cast_fp16")];
            tensor<int32, [4]> var_1903_perm_0 = const()[name = string("op_1903_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1905 = const()[name = string("op_1905"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_1903_cast_fp16 = transpose(perm = var_1903_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [4, 64, 896]> var_1906_cast_fp16 = reshape(shape = var_1905, x = var_1903_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301852608))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302655488))))[name = string("lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_80_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = var_1906_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_45_cast_fp16 = add(x = x_43_cast_fp16, y = linear_80_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1912_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_1912_cast_fp16")];
            tensor<int32, [1]> var_1914_axes_0 = const()[name = string("op_1914_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1914_cast_fp16 = reduce_max(axes = var_1914_axes_0, keep_dims = var_12, x = var_1912_cast_fp16)[name = string("op_1914_cast_fp16")];
            fp16 const_225_to_fp16 = const()[name = string("const_225_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_23_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_225_to_fp16, x = var_1914_cast_fp16)[name = string("clip_23_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = clip_23_cast_fp16)[name = string("xn_47_cast_fp16")];
            fp16 var_23_promoted_23_to_fp16 = const()[name = string("op_23_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1917_cast_fp16 = pow(x = xn_47_cast_fp16, y = var_23_promoted_23_to_fp16)[name = string("op_1917_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1917_cast_fp16)[name = string("var_47_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1920_cast_fp16 = mul(x = clip_23_cast_fp16, y = clip_23_cast_fp16)[name = string("op_1920_cast_fp16")];
            fp32 var_1921_epsilon_0 = const()[name = string("op_1921_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1921_cast_fp16 = inverse(epsilon = var_1921_epsilon_0, x = var_1920_cast_fp16)[name = string("op_1921_cast_fp16")];
            fp16 var_1922_to_fp16 = const()[name = string("op_1922_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1923_cast_fp16 = mul(x = var_1921_cast_fp16, y = var_1922_to_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1924_cast_fp16 = add(x = var_47_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1924_cast_fp16")];
            fp32 var_1925_epsilon_0 = const()[name = string("op_1925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1925_cast_fp16 = rsqrt(epsilon = var_1925_epsilon_0, x = var_1924_cast_fp16)[name = string("op_1925_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_47_cast_fp16 = mul(x = xn_47_cast_fp16, y = var_1925_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [896]> lm_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302657344)))];
            tensor<fp16, [4, 64, 896]> input_91_cast_fp16 = mul(x = lm_layers_11_post_attention_layernorm_weight_to_fp16, y = out_47_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302659200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307017408))))[name = string("lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_1933_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307027200))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311385408))))[name = string("lm_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_95_cast_fp16 = mul(x = var_1933_cast_fp16, y = linear_82_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311395200))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753408))))[name = string("lm_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_47_cast_fp16 = add(x = x_45_cast_fp16, y = linear_83_cast_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_1945_cast_fp16 = abs(x = x_47_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<int32, [1]> var_1947_axes_0 = const()[name = string("op_1947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_1947_cast_fp16 = reduce_max(axes = var_1947_axes_0, keep_dims = var_12, x = var_1945_cast_fp16)[name = string("op_1947_cast_fp16")];
            fp16 const_226_to_fp16 = const()[name = string("const_226_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_24_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_226_to_fp16, x = var_1947_cast_fp16)[name = string("clip_24_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_49_cast_fp16 = real_div(x = x_47_cast_fp16, y = clip_24_cast_fp16)[name = string("xn_49_cast_fp16")];
            fp16 var_23_promoted_24_to_fp16 = const()[name = string("op_23_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_1950_cast_fp16 = pow(x = xn_49_cast_fp16, y = var_23_promoted_24_to_fp16)[name = string("op_1950_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1950_cast_fp16)[name = string("var_49_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1953_cast_fp16 = mul(x = clip_24_cast_fp16, y = clip_24_cast_fp16)[name = string("op_1953_cast_fp16")];
            fp32 var_1954_epsilon_0 = const()[name = string("op_1954_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_1954_cast_fp16 = inverse(epsilon = var_1954_epsilon_0, x = var_1953_cast_fp16)[name = string("op_1954_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_1956_cast_fp16 = mul(x = var_1954_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_1957_cast_fp16 = add(x = var_49_cast_fp16, y = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            fp32 var_1958_epsilon_0 = const()[name = string("op_1958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_1958_cast_fp16 = rsqrt(epsilon = var_1958_epsilon_0, x = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_49_cast_fp16 = mul(x = xn_49_cast_fp16, y = var_1958_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755264)))];
            tensor<fp16, [4, 64, 896]> hidden_states_169_cast_fp16 = mul(x = lm_layers_12_input_layernorm_weight_to_fp16, y = out_49_cast_fp16)[name = string("hidden_states_169_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757120))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316560000))))[name = string("lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316561856)))];
            tensor<fp16, [4, 64, 896]> linear_84_cast_fp16 = linear(bias = lm_layers_12_self_attn_q_proj_bias_to_fp16, weight = lm_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_1970 = const()[name = string("op_1970"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_1971_cast_fp16 = reshape(shape = var_1970, x = linear_84_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = string("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316563712))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678464))))[name = string("lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316678784)))];
            tensor<fp16, [4, 64, 128]> linear_85_cast_fp16 = linear(bias = lm_layers_12_self_attn_k_proj_bias_to_fp16, weight = lm_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1977_cast_fp16 = reshape(shape = var_1976, x = linear_85_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = string("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316679104))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316793856))))[name = string("lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794176)))];
            tensor<fp16, [4, 64, 128]> linear_86_cast_fp16 = linear(bias = lm_layers_12_self_attn_v_proj_bias_to_fp16, weight = lm_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_169_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_1983_cast_fp16 = reshape(shape = var_1982, x = linear_86_cast_fp16)[name = string("op_1983_cast_fp16")];
            tensor<int32, [4]> hidden_states_175_perm_0 = const()[name = string("hidden_states_175_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1971_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [4, 14, 64, 64]> var_1987_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_1987_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_25_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_25_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_231_promoted_to_fp16 = const()[name = string("const_231_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_1998_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_231_promoted_to_fp16)[name = string("op_1998_cast_fp16")];
            bool var_2000_interleave_0 = const()[name = string("op_2000_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2000_cast_fp16 = concat(axis = var_20, interleave = var_2000_interleave_0, values = (var_1998_cast_fp16, x1_49_cast_fp16))[name = string("op_2000_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2001_cast_fp16 = mul(x = var_2000_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_25_cast_fp16 = add(x = var_1987_cast_fp16, y = var_2001_cast_fp16)[name = string("query_25_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1977_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [4, 2, 64, 64]> var_2003_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_25_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_25_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_234_promoted_to_fp16 = const()[name = string("const_234_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2014_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_234_promoted_to_fp16)[name = string("op_2014_cast_fp16")];
            bool var_2016_interleave_0 = const()[name = string("op_2016_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2016_cast_fp16 = concat(axis = var_20, interleave = var_2016_interleave_0, values = (var_2014_cast_fp16, x1_51_cast_fp16))[name = string("op_2016_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2017_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_171_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2017_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<int32, [1]> var_2026_axes_0 = const()[name = string("op_2026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2026_cast_fp16 = expand_dims(axes = var_2026_axes_0, x = hidden_states_171_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<int32, [5]> hidden_states_173_reps_0 = const()[name = string("hidden_states_173_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_173_cast_fp16 = tile(reps = hidden_states_173_reps_0, x = var_2026_cast_fp16)[name = string("hidden_states_173_cast_fp16")];
            tensor<int32, [4]> var_2034 = const()[name = string("op_2034"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_25_cast_fp16 = reshape(shape = var_2034, x = hidden_states_173_cast_fp16)[name = string("key_25_cast_fp16")];
            tensor<int32, [1]> var_2043_axes_0 = const()[name = string("op_2043_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_175_cast_fp16 = transpose(perm = hidden_states_175_perm_0, x = var_1983_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2043_cast_fp16 = expand_dims(axes = var_2043_axes_0, x = hidden_states_175_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [5]> hidden_states_177_reps_0 = const()[name = string("hidden_states_177_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_177_cast_fp16 = tile(reps = hidden_states_177_reps_0, x = var_2043_cast_fp16)[name = string("hidden_states_177_cast_fp16")];
            tensor<int32, [4]> var_2051 = const()[name = string("op_2051"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_25_cast_fp16 = reshape(shape = var_2051, x = hidden_states_177_cast_fp16)[name = string("value_25_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_25_cast_fp16 = mul(x = query_25_cast_fp16, y = var_13_to_fp16)[name = string("mul_25_cast_fp16")];
            bool matmul_12_transpose_y_0 = const()[name = string("matmul_12_transpose_y_0"), val = bool(true)];
            bool matmul_12_transpose_x_0 = const()[name = string("matmul_12_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_0, transpose_y = matmul_12_transpose_y_0, x = mul_25_cast_fp16, y = key_25_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_12_cast_fp16 = add(x = matmul_12_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_12_cast_fp16")];
            int32 softmax_12_axis_0 = const()[name = string("softmax_12_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_12_cast_fp16 = softmax(axis = softmax_12_axis_0, x = add_12_cast_fp16)[name = string("softmax_12_cast_fp16")];
            bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)];
            bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_12_cast_fp16, y = value_25_cast_fp16)[name = string("attn_output_49_cast_fp16")];
            tensor<int32, [4]> var_2054_perm_0 = const()[name = string("op_2054_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2056 = const()[name = string("op_2056"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2054_cast_fp16 = transpose(perm = var_2054_perm_0, x = attn_output_49_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [4, 64, 896]> var_2057_cast_fp16 = reshape(shape = var_2056, x = var_2054_cast_fp16)[name = string("op_2057_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316794496))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317597376))))[name = string("lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = var_2057_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_87_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2063_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2063_cast_fp16")];
            tensor<int32, [1]> var_2065_axes_0 = const()[name = string("op_2065_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2065_cast_fp16 = reduce_max(axes = var_2065_axes_0, keep_dims = var_12, x = var_2063_cast_fp16)[name = string("op_2065_cast_fp16")];
            fp16 const_243_to_fp16 = const()[name = string("const_243_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_25_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_243_to_fp16, x = var_2065_cast_fp16)[name = string("clip_25_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = clip_25_cast_fp16)[name = string("xn_51_cast_fp16")];
            fp16 var_23_promoted_25_to_fp16 = const()[name = string("op_23_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2068_cast_fp16 = pow(x = xn_51_cast_fp16, y = var_23_promoted_25_to_fp16)[name = string("op_2068_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_2068_cast_fp16)[name = string("var_51_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2071_cast_fp16 = mul(x = clip_25_cast_fp16, y = clip_25_cast_fp16)[name = string("op_2071_cast_fp16")];
            fp32 var_2072_epsilon_0 = const()[name = string("op_2072_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2072_cast_fp16 = inverse(epsilon = var_2072_epsilon_0, x = var_2071_cast_fp16)[name = string("op_2072_cast_fp16")];
            fp16 var_2073_to_fp16 = const()[name = string("op_2073_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2074_cast_fp16 = mul(x = var_2072_cast_fp16, y = var_2073_to_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2075_cast_fp16 = add(x = var_51_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2075_cast_fp16")];
            fp32 var_2076_epsilon_0 = const()[name = string("op_2076_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2076_cast_fp16 = rsqrt(epsilon = var_2076_epsilon_0, x = var_2075_cast_fp16)[name = string("op_2076_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_51_cast_fp16 = mul(x = xn_51_cast_fp16, y = var_2076_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [896]> lm_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317599232)))];
            tensor<fp16, [4, 64, 896]> input_99_cast_fp16 = mul(x = lm_layers_12_post_attention_layernorm_weight_to_fp16, y = out_51_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317601088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321959296))))[name = string("lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2084_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2084_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321969088))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326327296))))[name = string("lm_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_103_cast_fp16 = mul(x = var_2084_cast_fp16, y = linear_89_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326337088))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330695296))))[name = string("lm_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_90_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_51_cast_fp16 = add(x = x_49_cast_fp16, y = linear_90_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2096_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2096_cast_fp16")];
            tensor<int32, [1]> var_2098_axes_0 = const()[name = string("op_2098_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2098_cast_fp16 = reduce_max(axes = var_2098_axes_0, keep_dims = var_12, x = var_2096_cast_fp16)[name = string("op_2098_cast_fp16")];
            fp16 const_244_to_fp16 = const()[name = string("const_244_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_26_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_244_to_fp16, x = var_2098_cast_fp16)[name = string("clip_26_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_53_cast_fp16 = real_div(x = x_51_cast_fp16, y = clip_26_cast_fp16)[name = string("xn_53_cast_fp16")];
            fp16 var_23_promoted_26_to_fp16 = const()[name = string("op_23_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2101_cast_fp16 = pow(x = xn_53_cast_fp16, y = var_23_promoted_26_to_fp16)[name = string("op_2101_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_2101_cast_fp16)[name = string("var_53_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2104_cast_fp16 = mul(x = clip_26_cast_fp16, y = clip_26_cast_fp16)[name = string("op_2104_cast_fp16")];
            fp32 var_2105_epsilon_0 = const()[name = string("op_2105_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2105_cast_fp16 = inverse(epsilon = var_2105_epsilon_0, x = var_2104_cast_fp16)[name = string("op_2105_cast_fp16")];
            fp16 var_2106_to_fp16 = const()[name = string("op_2106_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2107_cast_fp16 = mul(x = var_2105_cast_fp16, y = var_2106_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2108_cast_fp16 = add(x = var_53_cast_fp16, y = var_2107_cast_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_53_cast_fp16 = mul(x = xn_53_cast_fp16, y = var_2109_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330697152)))];
            tensor<fp16, [4, 64, 896]> hidden_states_183_cast_fp16 = mul(x = lm_layers_13_input_layernorm_weight_to_fp16, y = out_53_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330699008))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331501888))))[name = string("lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331503744)))];
            tensor<fp16, [4, 64, 896]> linear_91_cast_fp16 = linear(bias = lm_layers_13_self_attn_q_proj_bias_to_fp16, weight = lm_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2122_cast_fp16 = reshape(shape = var_2121, x = linear_91_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331505600))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620352))))[name = string("lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620672)))];
            tensor<fp16, [4, 64, 128]> linear_92_cast_fp16 = linear(bias = lm_layers_13_self_attn_k_proj_bias_to_fp16, weight = lm_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2127 = const()[name = string("op_2127"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2128_cast_fp16 = reshape(shape = var_2127, x = linear_92_cast_fp16)[name = string("op_2128_cast_fp16")];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331620992))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331735744))))[name = string("lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736064)))];
            tensor<fp16, [4, 64, 128]> linear_93_cast_fp16 = linear(bias = lm_layers_13_self_attn_v_proj_bias_to_fp16, weight = lm_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_183_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2134_cast_fp16 = reshape(shape = var_2133, x = linear_93_cast_fp16)[name = string("op_2134_cast_fp16")];
            tensor<int32, [4]> hidden_states_189_perm_0 = const()[name = string("hidden_states_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_2122_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [4, 14, 64, 64]> var_2138_cast_fp16 = mul(x = q_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2138_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_27_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_27_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_249_promoted_to_fp16 = const()[name = string("const_249_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2149_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_249_promoted_to_fp16)[name = string("op_2149_cast_fp16")];
            bool var_2151_interleave_0 = const()[name = string("op_2151_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2151_cast_fp16 = concat(axis = var_20, interleave = var_2151_interleave_0, values = (var_2149_cast_fp16, x1_53_cast_fp16))[name = string("op_2151_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2152_cast_fp16 = mul(x = var_2151_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_27_cast_fp16 = add(x = var_2138_cast_fp16, y = var_2152_cast_fp16)[name = string("query_27_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_2128_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [4, 2, 64, 64]> var_2154_cast_fp16 = mul(x = k_27_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_27_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_27_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_252_promoted_to_fp16 = const()[name = string("const_252_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2165_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_252_promoted_to_fp16)[name = string("op_2165_cast_fp16")];
            bool var_2167_interleave_0 = const()[name = string("op_2167_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2167_cast_fp16 = concat(axis = var_20, interleave = var_2167_interleave_0, values = (var_2165_cast_fp16, x1_55_cast_fp16))[name = string("op_2167_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2168_cast_fp16 = mul(x = var_2167_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2168_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_185_cast_fp16 = add(x = var_2154_cast_fp16, y = var_2168_cast_fp16)[name = string("hidden_states_185_cast_fp16")];
            tensor<int32, [1]> var_2177_axes_0 = const()[name = string("op_2177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2177_cast_fp16 = expand_dims(axes = var_2177_axes_0, x = hidden_states_185_cast_fp16)[name = string("op_2177_cast_fp16")];
            tensor<int32, [5]> hidden_states_187_reps_0 = const()[name = string("hidden_states_187_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_187_cast_fp16 = tile(reps = hidden_states_187_reps_0, x = var_2177_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<int32, [4]> var_2185 = const()[name = string("op_2185"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_27_cast_fp16 = reshape(shape = var_2185, x = hidden_states_187_cast_fp16)[name = string("key_27_cast_fp16")];
            tensor<int32, [1]> var_2194_axes_0 = const()[name = string("op_2194_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_189_cast_fp16 = transpose(perm = hidden_states_189_perm_0, x = var_2134_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2194_cast_fp16 = expand_dims(axes = var_2194_axes_0, x = hidden_states_189_cast_fp16)[name = string("op_2194_cast_fp16")];
            tensor<int32, [5]> hidden_states_191_reps_0 = const()[name = string("hidden_states_191_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_191_cast_fp16 = tile(reps = hidden_states_191_reps_0, x = var_2194_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_27_cast_fp16 = reshape(shape = var_2202, x = hidden_states_191_cast_fp16)[name = string("value_27_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_27_cast_fp16 = mul(x = query_27_cast_fp16, y = var_13_to_fp16)[name = string("mul_27_cast_fp16")];
            bool matmul_13_transpose_y_0 = const()[name = string("matmul_13_transpose_y_0"), val = bool(true)];
            bool matmul_13_transpose_x_0 = const()[name = string("matmul_13_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_0, transpose_y = matmul_13_transpose_y_0, x = mul_27_cast_fp16, y = key_27_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_13_cast_fp16 = add(x = matmul_13_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_13_cast_fp16")];
            int32 softmax_13_axis_0 = const()[name = string("softmax_13_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_13_cast_fp16 = softmax(axis = softmax_13_axis_0, x = add_13_cast_fp16)[name = string("softmax_13_cast_fp16")];
            bool attn_output_53_transpose_x_0 = const()[name = string("attn_output_53_transpose_x_0"), val = bool(false)];
            bool attn_output_53_transpose_y_0 = const()[name = string("attn_output_53_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_53_cast_fp16 = matmul(transpose_x = attn_output_53_transpose_x_0, transpose_y = attn_output_53_transpose_y_0, x = softmax_13_cast_fp16, y = value_27_cast_fp16)[name = string("attn_output_53_cast_fp16")];
            tensor<int32, [4]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2207 = const()[name = string("op_2207"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = attn_output_53_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [4, 64, 896]> var_2208_cast_fp16 = reshape(shape = var_2207, x = var_2205_cast_fp16)[name = string("op_2208_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331736384))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332539264))))[name = string("lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_94_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = var_2208_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_53_cast_fp16 = add(x = x_51_cast_fp16, y = linear_94_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2214_cast_fp16 = abs(x = x_53_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<int32, [1]> var_2216_axes_0 = const()[name = string("op_2216_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2216_cast_fp16 = reduce_max(axes = var_2216_axes_0, keep_dims = var_12, x = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_27_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_261_to_fp16, x = var_2216_cast_fp16)[name = string("clip_27_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_55_cast_fp16 = real_div(x = x_53_cast_fp16, y = clip_27_cast_fp16)[name = string("xn_55_cast_fp16")];
            fp16 var_23_promoted_27_to_fp16 = const()[name = string("op_23_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2219_cast_fp16 = pow(x = xn_55_cast_fp16, y = var_23_promoted_27_to_fp16)[name = string("op_2219_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_2219_cast_fp16)[name = string("var_55_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2222_cast_fp16 = mul(x = clip_27_cast_fp16, y = clip_27_cast_fp16)[name = string("op_2222_cast_fp16")];
            fp32 var_2223_epsilon_0 = const()[name = string("op_2223_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2223_cast_fp16 = inverse(epsilon = var_2223_epsilon_0, x = var_2222_cast_fp16)[name = string("op_2223_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2225_cast_fp16 = mul(x = var_2223_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2226_cast_fp16 = add(x = var_55_cast_fp16, y = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp32 var_2227_epsilon_0 = const()[name = string("op_2227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2227_cast_fp16 = rsqrt(epsilon = var_2227_epsilon_0, x = var_2226_cast_fp16)[name = string("op_2227_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_55_cast_fp16 = mul(x = xn_55_cast_fp16, y = var_2227_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [896]> lm_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332541120)))];
            tensor<fp16, [4, 64, 896]> input_107_cast_fp16 = mul(x = lm_layers_13_post_attention_layernorm_weight_to_fp16, y = out_55_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332542976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336901184))))[name = string("lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2235_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336910976))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341269184))))[name = string("lm_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_111_cast_fp16 = mul(x = var_2235_cast_fp16, y = linear_96_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341278976))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345637184))))[name = string("lm_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_97_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_97_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2247_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2247_cast_fp16")];
            tensor<int32, [1]> var_2249_axes_0 = const()[name = string("op_2249_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2249_cast_fp16 = reduce_max(axes = var_2249_axes_0, keep_dims = var_12, x = var_2247_cast_fp16)[name = string("op_2249_cast_fp16")];
            fp16 const_262_to_fp16 = const()[name = string("const_262_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_28_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_262_to_fp16, x = var_2249_cast_fp16)[name = string("clip_28_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_57_cast_fp16 = real_div(x = x_55_cast_fp16, y = clip_28_cast_fp16)[name = string("xn_57_cast_fp16")];
            fp16 var_23_promoted_28_to_fp16 = const()[name = string("op_23_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2252_cast_fp16 = pow(x = xn_57_cast_fp16, y = var_23_promoted_28_to_fp16)[name = string("op_2252_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_2252_cast_fp16)[name = string("var_57_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2255_cast_fp16 = mul(x = clip_28_cast_fp16, y = clip_28_cast_fp16)[name = string("op_2255_cast_fp16")];
            fp32 var_2256_epsilon_0 = const()[name = string("op_2256_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2256_cast_fp16 = inverse(epsilon = var_2256_epsilon_0, x = var_2255_cast_fp16)[name = string("op_2256_cast_fp16")];
            fp16 var_2257_to_fp16 = const()[name = string("op_2257_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2258_cast_fp16 = mul(x = var_2256_cast_fp16, y = var_2257_to_fp16)[name = string("op_2258_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2259_cast_fp16 = add(x = var_57_cast_fp16, y = var_2258_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_57_cast_fp16 = mul(x = xn_57_cast_fp16, y = var_2260_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345639040)))];
            tensor<fp16, [4, 64, 896]> hidden_states_197_cast_fp16 = mul(x = lm_layers_14_input_layernorm_weight_to_fp16, y = out_57_cast_fp16)[name = string("hidden_states_197_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345640896))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346443776))))[name = string("lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346445632)))];
            tensor<fp16, [4, 64, 896]> linear_98_cast_fp16 = linear(bias = lm_layers_14_self_attn_q_proj_bias_to_fp16, weight = lm_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2273_cast_fp16 = reshape(shape = var_2272, x = linear_98_cast_fp16)[name = string("op_2273_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = string("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346447488))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562240))))[name = string("lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562560)))];
            tensor<fp16, [4, 64, 128]> linear_99_cast_fp16 = linear(bias = lm_layers_14_self_attn_k_proj_bias_to_fp16, weight = lm_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2278 = const()[name = string("op_2278"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2279_cast_fp16 = reshape(shape = var_2278, x = linear_99_cast_fp16)[name = string("op_2279_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346562880))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677632))))[name = string("lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346677952)))];
            tensor<fp16, [4, 64, 128]> linear_100_cast_fp16 = linear(bias = lm_layers_14_self_attn_v_proj_bias_to_fp16, weight = lm_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_197_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2285_cast_fp16 = reshape(shape = var_2284, x = linear_100_cast_fp16)[name = string("op_2285_cast_fp16")];
            tensor<int32, [4]> hidden_states_203_perm_0 = const()[name = string("hidden_states_203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_2273_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [4, 14, 64, 64]> var_2289_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_29_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_29_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_267_promoted_to_fp16 = const()[name = string("const_267_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2300_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_267_promoted_to_fp16)[name = string("op_2300_cast_fp16")];
            bool var_2302_interleave_0 = const()[name = string("op_2302_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2302_cast_fp16 = concat(axis = var_20, interleave = var_2302_interleave_0, values = (var_2300_cast_fp16, x1_57_cast_fp16))[name = string("op_2302_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2303_cast_fp16 = mul(x = var_2302_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2303_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_29_cast_fp16 = add(x = var_2289_cast_fp16, y = var_2303_cast_fp16)[name = string("query_29_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_2279_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [4, 2, 64, 64]> var_2305_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2305_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_29_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_29_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_270_promoted_to_fp16 = const()[name = string("const_270_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2316_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_270_promoted_to_fp16)[name = string("op_2316_cast_fp16")];
            bool var_2318_interleave_0 = const()[name = string("op_2318_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2318_cast_fp16 = concat(axis = var_20, interleave = var_2318_interleave_0, values = (var_2316_cast_fp16, x1_59_cast_fp16))[name = string("op_2318_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_199_cast_fp16 = add(x = var_2305_cast_fp16, y = var_2319_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            tensor<int32, [1]> var_2328_axes_0 = const()[name = string("op_2328_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2328_cast_fp16 = expand_dims(axes = var_2328_axes_0, x = hidden_states_199_cast_fp16)[name = string("op_2328_cast_fp16")];
            tensor<int32, [5]> hidden_states_201_reps_0 = const()[name = string("hidden_states_201_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_201_cast_fp16 = tile(reps = hidden_states_201_reps_0, x = var_2328_cast_fp16)[name = string("hidden_states_201_cast_fp16")];
            tensor<int32, [4]> var_2336 = const()[name = string("op_2336"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_29_cast_fp16 = reshape(shape = var_2336, x = hidden_states_201_cast_fp16)[name = string("key_29_cast_fp16")];
            tensor<int32, [1]> var_2345_axes_0 = const()[name = string("op_2345_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_203_cast_fp16 = transpose(perm = hidden_states_203_perm_0, x = var_2285_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2345_cast_fp16 = expand_dims(axes = var_2345_axes_0, x = hidden_states_203_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<int32, [5]> hidden_states_205_reps_0 = const()[name = string("hidden_states_205_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_205_cast_fp16 = tile(reps = hidden_states_205_reps_0, x = var_2345_cast_fp16)[name = string("hidden_states_205_cast_fp16")];
            tensor<int32, [4]> var_2353 = const()[name = string("op_2353"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_29_cast_fp16 = reshape(shape = var_2353, x = hidden_states_205_cast_fp16)[name = string("value_29_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_29_cast_fp16 = mul(x = query_29_cast_fp16, y = var_13_to_fp16)[name = string("mul_29_cast_fp16")];
            bool matmul_14_transpose_y_0 = const()[name = string("matmul_14_transpose_y_0"), val = bool(true)];
            bool matmul_14_transpose_x_0 = const()[name = string("matmul_14_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_0, transpose_y = matmul_14_transpose_y_0, x = mul_29_cast_fp16, y = key_29_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_14_cast_fp16 = add(x = matmul_14_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_14_cast_fp16")];
            int32 softmax_14_axis_0 = const()[name = string("softmax_14_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_14_cast_fp16 = softmax(axis = softmax_14_axis_0, x = add_14_cast_fp16)[name = string("softmax_14_cast_fp16")];
            bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)];
            bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_14_cast_fp16, y = value_29_cast_fp16)[name = string("attn_output_57_cast_fp16")];
            tensor<int32, [4]> var_2356_perm_0 = const()[name = string("op_2356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2358 = const()[name = string("op_2358"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2356_cast_fp16 = transpose(perm = var_2356_perm_0, x = attn_output_57_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [4, 64, 896]> var_2359_cast_fp16 = reshape(shape = var_2358, x = var_2356_cast_fp16)[name = string("op_2359_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346678272))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347481152))))[name = string("lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_101_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = var_2359_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_57_cast_fp16 = add(x = x_55_cast_fp16, y = linear_101_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2365_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2365_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2367_cast_fp16 = reduce_max(axes = var_2367_axes_0, keep_dims = var_12, x = var_2365_cast_fp16)[name = string("op_2367_cast_fp16")];
            fp16 const_279_to_fp16 = const()[name = string("const_279_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_29_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_279_to_fp16, x = var_2367_cast_fp16)[name = string("clip_29_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = clip_29_cast_fp16)[name = string("xn_59_cast_fp16")];
            fp16 var_23_promoted_29_to_fp16 = const()[name = string("op_23_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2370_cast_fp16 = pow(x = xn_59_cast_fp16, y = var_23_promoted_29_to_fp16)[name = string("op_2370_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_2370_cast_fp16)[name = string("var_59_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2373_cast_fp16 = mul(x = clip_29_cast_fp16, y = clip_29_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp32 var_2374_epsilon_0 = const()[name = string("op_2374_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2374_cast_fp16 = inverse(epsilon = var_2374_epsilon_0, x = var_2373_cast_fp16)[name = string("op_2374_cast_fp16")];
            fp16 var_2375_to_fp16 = const()[name = string("op_2375_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2376_cast_fp16 = mul(x = var_2374_cast_fp16, y = var_2375_to_fp16)[name = string("op_2376_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2377_cast_fp16 = add(x = var_59_cast_fp16, y = var_2376_cast_fp16)[name = string("op_2377_cast_fp16")];
            fp32 var_2378_epsilon_0 = const()[name = string("op_2378_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2378_cast_fp16 = rsqrt(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = string("op_2378_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_59_cast_fp16 = mul(x = xn_59_cast_fp16, y = var_2378_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [896]> lm_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347483008)))];
            tensor<fp16, [4, 64, 896]> input_115_cast_fp16 = mul(x = lm_layers_14_post_attention_layernorm_weight_to_fp16, y = out_59_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347484864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351843072))))[name = string("lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2386_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_2386_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351852864))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356211072))))[name = string("lm_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_119_cast_fp16 = mul(x = var_2386_cast_fp16, y = linear_103_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356220864))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360579072))))[name = string("lm_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_104_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_59_cast_fp16 = add(x = x_57_cast_fp16, y = linear_104_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2398_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2398_cast_fp16")];
            tensor<int32, [1]> var_2400_axes_0 = const()[name = string("op_2400_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2400_cast_fp16 = reduce_max(axes = var_2400_axes_0, keep_dims = var_12, x = var_2398_cast_fp16)[name = string("op_2400_cast_fp16")];
            fp16 const_280_to_fp16 = const()[name = string("const_280_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_30_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_280_to_fp16, x = var_2400_cast_fp16)[name = string("clip_30_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_61_cast_fp16 = real_div(x = x_59_cast_fp16, y = clip_30_cast_fp16)[name = string("xn_61_cast_fp16")];
            fp16 var_23_promoted_30_to_fp16 = const()[name = string("op_23_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2403_cast_fp16 = pow(x = xn_61_cast_fp16, y = var_23_promoted_30_to_fp16)[name = string("op_2403_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_2403_cast_fp16)[name = string("var_61_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2406_cast_fp16 = mul(x = clip_30_cast_fp16, y = clip_30_cast_fp16)[name = string("op_2406_cast_fp16")];
            fp32 var_2407_epsilon_0 = const()[name = string("op_2407_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2407_cast_fp16 = inverse(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16)[name = string("op_2407_cast_fp16")];
            fp16 var_2408_to_fp16 = const()[name = string("op_2408_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2409_cast_fp16 = mul(x = var_2407_cast_fp16, y = var_2408_to_fp16)[name = string("op_2409_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2410_cast_fp16 = add(x = var_61_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2410_cast_fp16")];
            fp32 var_2411_epsilon_0 = const()[name = string("op_2411_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2411_cast_fp16 = rsqrt(epsilon = var_2411_epsilon_0, x = var_2410_cast_fp16)[name = string("op_2411_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_61_cast_fp16 = mul(x = xn_61_cast_fp16, y = var_2411_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360580928)))];
            tensor<fp16, [4, 64, 896]> hidden_states_211_cast_fp16 = mul(x = lm_layers_15_input_layernorm_weight_to_fp16, y = out_61_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360582784))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361385664))))[name = string("lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361387520)))];
            tensor<fp16, [4, 64, 896]> linear_105_cast_fp16 = linear(bias = lm_layers_15_self_attn_q_proj_bias_to_fp16, weight = lm_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2424_cast_fp16 = reshape(shape = var_2423, x = linear_105_cast_fp16)[name = string("op_2424_cast_fp16")];
            tensor<int32, [4]> q_31_perm_0 = const()[name = string("q_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361389376))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504128))))[name = string("lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504448)))];
            tensor<fp16, [4, 64, 128]> linear_106_cast_fp16 = linear(bias = lm_layers_15_self_attn_k_proj_bias_to_fp16, weight = lm_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = string("op_2429"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_106_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361504768))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619520))))[name = string("lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361619840)))];
            tensor<fp16, [4, 64, 128]> linear_107_cast_fp16 = linear(bias = lm_layers_15_self_attn_v_proj_bias_to_fp16, weight = lm_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_211_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2435 = const()[name = string("op_2435"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2436_cast_fp16 = reshape(shape = var_2435, x = linear_107_cast_fp16)[name = string("op_2436_cast_fp16")];
            tensor<int32, [4]> hidden_states_217_perm_0 = const()[name = string("hidden_states_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_31_cast_fp16 = transpose(perm = q_31_perm_0, x = var_2424_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [4, 14, 64, 64]> var_2440_cast_fp16 = mul(x = q_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_31_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_31_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_285_promoted_to_fp16 = const()[name = string("const_285_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2451_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_285_promoted_to_fp16)[name = string("op_2451_cast_fp16")];
            bool var_2453_interleave_0 = const()[name = string("op_2453_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2453_cast_fp16 = concat(axis = var_20, interleave = var_2453_interleave_0, values = (var_2451_cast_fp16, x1_61_cast_fp16))[name = string("op_2453_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2454_cast_fp16 = mul(x = var_2453_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2454_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_31_cast_fp16 = add(x = var_2440_cast_fp16, y = var_2454_cast_fp16)[name = string("query_31_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = var_2430_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [4, 2, 64, 64]> var_2456_cast_fp16 = mul(x = k_31_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2456_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_31_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_31_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_288_promoted_to_fp16 = const()[name = string("const_288_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2467_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_288_promoted_to_fp16)[name = string("op_2467_cast_fp16")];
            bool var_2469_interleave_0 = const()[name = string("op_2469_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2469_cast_fp16 = concat(axis = var_20, interleave = var_2469_interleave_0, values = (var_2467_cast_fp16, x1_63_cast_fp16))[name = string("op_2469_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2470_cast_fp16 = mul(x = var_2469_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2470_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_213_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2470_cast_fp16)[name = string("hidden_states_213_cast_fp16")];
            tensor<int32, [1]> var_2479_axes_0 = const()[name = string("op_2479_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2479_cast_fp16 = expand_dims(axes = var_2479_axes_0, x = hidden_states_213_cast_fp16)[name = string("op_2479_cast_fp16")];
            tensor<int32, [5]> hidden_states_215_reps_0 = const()[name = string("hidden_states_215_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_215_cast_fp16 = tile(reps = hidden_states_215_reps_0, x = var_2479_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            tensor<int32, [4]> var_2487 = const()[name = string("op_2487"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_31_cast_fp16 = reshape(shape = var_2487, x = hidden_states_215_cast_fp16)[name = string("key_31_cast_fp16")];
            tensor<int32, [1]> var_2496_axes_0 = const()[name = string("op_2496_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_217_cast_fp16 = transpose(perm = hidden_states_217_perm_0, x = var_2436_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2496_cast_fp16 = expand_dims(axes = var_2496_axes_0, x = hidden_states_217_cast_fp16)[name = string("op_2496_cast_fp16")];
            tensor<int32, [5]> hidden_states_219_reps_0 = const()[name = string("hidden_states_219_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_219_cast_fp16 = tile(reps = hidden_states_219_reps_0, x = var_2496_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<int32, [4]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_31_cast_fp16 = reshape(shape = var_2504, x = hidden_states_219_cast_fp16)[name = string("value_31_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_31_cast_fp16 = mul(x = query_31_cast_fp16, y = var_13_to_fp16)[name = string("mul_31_cast_fp16")];
            bool matmul_15_transpose_y_0 = const()[name = string("matmul_15_transpose_y_0"), val = bool(true)];
            bool matmul_15_transpose_x_0 = const()[name = string("matmul_15_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_0, transpose_y = matmul_15_transpose_y_0, x = mul_31_cast_fp16, y = key_31_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_15_cast_fp16 = add(x = matmul_15_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_15_cast_fp16")];
            int32 softmax_15_axis_0 = const()[name = string("softmax_15_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_15_cast_fp16 = softmax(axis = softmax_15_axis_0, x = add_15_cast_fp16)[name = string("softmax_15_cast_fp16")];
            bool attn_output_61_transpose_x_0 = const()[name = string("attn_output_61_transpose_x_0"), val = bool(false)];
            bool attn_output_61_transpose_y_0 = const()[name = string("attn_output_61_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_15_cast_fp16, y = value_31_cast_fp16)[name = string("attn_output_61_cast_fp16")];
            tensor<int32, [4]> var_2507_perm_0 = const()[name = string("op_2507_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2509 = const()[name = string("op_2509"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2507_cast_fp16 = transpose(perm = var_2507_perm_0, x = attn_output_61_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [4, 64, 896]> var_2510_cast_fp16 = reshape(shape = var_2509, x = var_2507_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361620160))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362423040))))[name = string("lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_108_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = var_2510_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_108_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2516_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2516_cast_fp16")];
            tensor<int32, [1]> var_2518_axes_0 = const()[name = string("op_2518_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2518_cast_fp16 = reduce_max(axes = var_2518_axes_0, keep_dims = var_12, x = var_2516_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 const_297_to_fp16 = const()[name = string("const_297_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_31_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_297_to_fp16, x = var_2518_cast_fp16)[name = string("clip_31_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = clip_31_cast_fp16)[name = string("xn_63_cast_fp16")];
            fp16 var_23_promoted_31_to_fp16 = const()[name = string("op_23_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2521_cast_fp16 = pow(x = xn_63_cast_fp16, y = var_23_promoted_31_to_fp16)[name = string("op_2521_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_2521_cast_fp16)[name = string("var_63_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2524_cast_fp16 = mul(x = clip_31_cast_fp16, y = clip_31_cast_fp16)[name = string("op_2524_cast_fp16")];
            fp32 var_2525_epsilon_0 = const()[name = string("op_2525_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2525_cast_fp16 = inverse(epsilon = var_2525_epsilon_0, x = var_2524_cast_fp16)[name = string("op_2525_cast_fp16")];
            fp16 var_2526_to_fp16 = const()[name = string("op_2526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2527_cast_fp16 = mul(x = var_2525_cast_fp16, y = var_2526_to_fp16)[name = string("op_2527_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2528_cast_fp16 = add(x = var_63_cast_fp16, y = var_2527_cast_fp16)[name = string("op_2528_cast_fp16")];
            fp32 var_2529_epsilon_0 = const()[name = string("op_2529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2529_cast_fp16 = rsqrt(epsilon = var_2529_epsilon_0, x = var_2528_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_63_cast_fp16 = mul(x = xn_63_cast_fp16, y = var_2529_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [896]> lm_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362424896)))];
            tensor<fp16, [4, 64, 896]> input_123_cast_fp16 = mul(x = lm_layers_15_post_attention_layernorm_weight_to_fp16, y = out_63_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362426752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366784960))))[name = string("lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2537_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_2537_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(366794752))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371152960))))[name = string("lm_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_127_cast_fp16 = mul(x = var_2537_cast_fp16, y = linear_110_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371162752))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375520960))))[name = string("lm_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_111_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_63_cast_fp16 = add(x = x_61_cast_fp16, y = linear_111_cast_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2549_cast_fp16 = abs(x = x_63_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [1]> var_2551_axes_0 = const()[name = string("op_2551_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2551_cast_fp16 = reduce_max(axes = var_2551_axes_0, keep_dims = var_12, x = var_2549_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 const_298_to_fp16 = const()[name = string("const_298_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_32_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_298_to_fp16, x = var_2551_cast_fp16)[name = string("clip_32_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_65_cast_fp16 = real_div(x = x_63_cast_fp16, y = clip_32_cast_fp16)[name = string("xn_65_cast_fp16")];
            fp16 var_23_promoted_32_to_fp16 = const()[name = string("op_23_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2554_cast_fp16 = pow(x = xn_65_cast_fp16, y = var_23_promoted_32_to_fp16)[name = string("op_2554_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_2554_cast_fp16)[name = string("var_65_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2557_cast_fp16 = mul(x = clip_32_cast_fp16, y = clip_32_cast_fp16)[name = string("op_2557_cast_fp16")];
            fp32 var_2558_epsilon_0 = const()[name = string("op_2558_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2558_cast_fp16 = inverse(epsilon = var_2558_epsilon_0, x = var_2557_cast_fp16)[name = string("op_2558_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2560_cast_fp16 = mul(x = var_2558_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2561_cast_fp16 = add(x = var_65_cast_fp16, y = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            fp32 var_2562_epsilon_0 = const()[name = string("op_2562_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2562_cast_fp16 = rsqrt(epsilon = var_2562_epsilon_0, x = var_2561_cast_fp16)[name = string("op_2562_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_65_cast_fp16 = mul(x = xn_65_cast_fp16, y = var_2562_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375522816)))];
            tensor<fp16, [4, 64, 896]> hidden_states_225_cast_fp16 = mul(x = lm_layers_16_input_layernorm_weight_to_fp16, y = out_65_cast_fp16)[name = string("hidden_states_225_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375524672))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376327552))))[name = string("lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376329408)))];
            tensor<fp16, [4, 64, 896]> linear_112_cast_fp16 = linear(bias = lm_layers_16_self_attn_q_proj_bias_to_fp16, weight = lm_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2575_cast_fp16 = reshape(shape = var_2574, x = linear_112_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376331264))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446016))))[name = string("lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446336)))];
            tensor<fp16, [4, 64, 128]> linear_113_cast_fp16 = linear(bias = lm_layers_16_self_attn_k_proj_bias_to_fp16, weight = lm_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2580 = const()[name = string("op_2580"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2581_cast_fp16 = reshape(shape = var_2580, x = linear_113_cast_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = string("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376446656))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561408))))[name = string("lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376561728)))];
            tensor<fp16, [4, 64, 128]> linear_114_cast_fp16 = linear(bias = lm_layers_16_self_attn_v_proj_bias_to_fp16, weight = lm_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_225_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2586 = const()[name = string("op_2586"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2587_cast_fp16 = reshape(shape = var_2586, x = linear_114_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<int32, [4]> hidden_states_231_perm_0 = const()[name = string("hidden_states_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_2575_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [4, 14, 64, 64]> var_2591_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2591_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_33_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_33_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_303_promoted_to_fp16 = const()[name = string("const_303_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2602_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_303_promoted_to_fp16)[name = string("op_2602_cast_fp16")];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2604_cast_fp16 = concat(axis = var_20, interleave = var_2604_interleave_0, values = (var_2602_cast_fp16, x1_65_cast_fp16))[name = string("op_2604_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2605_cast_fp16 = mul(x = var_2604_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2605_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_33_cast_fp16 = add(x = var_2591_cast_fp16, y = var_2605_cast_fp16)[name = string("query_33_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_2581_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [4, 2, 64, 64]> var_2607_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_33_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_33_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_306_promoted_to_fp16 = const()[name = string("const_306_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2618_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_306_promoted_to_fp16)[name = string("op_2618_cast_fp16")];
            bool var_2620_interleave_0 = const()[name = string("op_2620_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2620_cast_fp16 = concat(axis = var_20, interleave = var_2620_interleave_0, values = (var_2618_cast_fp16, x1_67_cast_fp16))[name = string("op_2620_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2621_cast_fp16 = mul(x = var_2620_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2621_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_227_cast_fp16 = add(x = var_2607_cast_fp16, y = var_2621_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2630_cast_fp16 = expand_dims(axes = var_2630_axes_0, x = hidden_states_227_cast_fp16)[name = string("op_2630_cast_fp16")];
            tensor<int32, [5]> hidden_states_229_reps_0 = const()[name = string("hidden_states_229_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_229_cast_fp16 = tile(reps = hidden_states_229_reps_0, x = var_2630_cast_fp16)[name = string("hidden_states_229_cast_fp16")];
            tensor<int32, [4]> var_2638 = const()[name = string("op_2638"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_33_cast_fp16 = reshape(shape = var_2638, x = hidden_states_229_cast_fp16)[name = string("key_33_cast_fp16")];
            tensor<int32, [1]> var_2647_axes_0 = const()[name = string("op_2647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_231_cast_fp16 = transpose(perm = hidden_states_231_perm_0, x = var_2587_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2647_cast_fp16 = expand_dims(axes = var_2647_axes_0, x = hidden_states_231_cast_fp16)[name = string("op_2647_cast_fp16")];
            tensor<int32, [5]> hidden_states_233_reps_0 = const()[name = string("hidden_states_233_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_233_cast_fp16 = tile(reps = hidden_states_233_reps_0, x = var_2647_cast_fp16)[name = string("hidden_states_233_cast_fp16")];
            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_33_cast_fp16 = reshape(shape = var_2655, x = hidden_states_233_cast_fp16)[name = string("value_33_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_33_cast_fp16 = mul(x = query_33_cast_fp16, y = var_13_to_fp16)[name = string("mul_33_cast_fp16")];
            bool matmul_16_transpose_y_0 = const()[name = string("matmul_16_transpose_y_0"), val = bool(true)];
            bool matmul_16_transpose_x_0 = const()[name = string("matmul_16_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_0, transpose_y = matmul_16_transpose_y_0, x = mul_33_cast_fp16, y = key_33_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_16_cast_fp16 = add(x = matmul_16_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_16_cast_fp16")];
            int32 softmax_16_axis_0 = const()[name = string("softmax_16_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_16_cast_fp16 = softmax(axis = softmax_16_axis_0, x = add_16_cast_fp16)[name = string("softmax_16_cast_fp16")];
            bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)];
            bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_16_cast_fp16, y = value_33_cast_fp16)[name = string("attn_output_65_cast_fp16")];
            tensor<int32, [4]> var_2658_perm_0 = const()[name = string("op_2658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2658_cast_fp16 = transpose(perm = var_2658_perm_0, x = attn_output_65_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [4, 64, 896]> var_2661_cast_fp16 = reshape(shape = var_2660, x = var_2658_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376562048))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377364928))))[name = string("lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_115_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = var_2661_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_65_cast_fp16 = add(x = x_63_cast_fp16, y = linear_115_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2667_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2667_cast_fp16")];
            tensor<int32, [1]> var_2669_axes_0 = const()[name = string("op_2669_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2669_cast_fp16 = reduce_max(axes = var_2669_axes_0, keep_dims = var_12, x = var_2667_cast_fp16)[name = string("op_2669_cast_fp16")];
            fp16 const_315_to_fp16 = const()[name = string("const_315_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_33_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_315_to_fp16, x = var_2669_cast_fp16)[name = string("clip_33_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = clip_33_cast_fp16)[name = string("xn_67_cast_fp16")];
            fp16 var_23_promoted_33_to_fp16 = const()[name = string("op_23_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2672_cast_fp16 = pow(x = xn_67_cast_fp16, y = var_23_promoted_33_to_fp16)[name = string("op_2672_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_2672_cast_fp16)[name = string("var_67_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2675_cast_fp16 = mul(x = clip_33_cast_fp16, y = clip_33_cast_fp16)[name = string("op_2675_cast_fp16")];
            fp32 var_2676_epsilon_0 = const()[name = string("op_2676_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2676_cast_fp16 = inverse(epsilon = var_2676_epsilon_0, x = var_2675_cast_fp16)[name = string("op_2676_cast_fp16")];
            fp16 var_2677_to_fp16 = const()[name = string("op_2677_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2678_cast_fp16 = mul(x = var_2676_cast_fp16, y = var_2677_to_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2679_cast_fp16 = add(x = var_67_cast_fp16, y = var_2678_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp32 var_2680_epsilon_0 = const()[name = string("op_2680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2680_cast_fp16 = rsqrt(epsilon = var_2680_epsilon_0, x = var_2679_cast_fp16)[name = string("op_2680_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_67_cast_fp16 = mul(x = xn_67_cast_fp16, y = var_2680_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [896]> lm_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377366784)))];
            tensor<fp16, [4, 64, 896]> input_131_cast_fp16 = mul(x = lm_layers_16_post_attention_layernorm_weight_to_fp16, y = out_67_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377368640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381726848))))[name = string("lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2688_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_2688_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(381736640))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386094848))))[name = string("lm_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_135_cast_fp16 = mul(x = var_2688_cast_fp16, y = linear_117_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386104640))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390462848))))[name = string("lm_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_118_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_118_cast_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2700_cast_fp16 = abs(x = x_67_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<int32, [1]> var_2702_axes_0 = const()[name = string("op_2702_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2702_cast_fp16 = reduce_max(axes = var_2702_axes_0, keep_dims = var_12, x = var_2700_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_316_to_fp16 = const()[name = string("const_316_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_34_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_316_to_fp16, x = var_2702_cast_fp16)[name = string("clip_34_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_69_cast_fp16 = real_div(x = x_67_cast_fp16, y = clip_34_cast_fp16)[name = string("xn_69_cast_fp16")];
            fp16 var_23_promoted_34_to_fp16 = const()[name = string("op_23_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2705_cast_fp16 = pow(x = xn_69_cast_fp16, y = var_23_promoted_34_to_fp16)[name = string("op_2705_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_2705_cast_fp16)[name = string("var_69_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2708_cast_fp16 = mul(x = clip_34_cast_fp16, y = clip_34_cast_fp16)[name = string("op_2708_cast_fp16")];
            fp32 var_2709_epsilon_0 = const()[name = string("op_2709_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2709_cast_fp16 = inverse(epsilon = var_2709_epsilon_0, x = var_2708_cast_fp16)[name = string("op_2709_cast_fp16")];
            fp16 var_2710_to_fp16 = const()[name = string("op_2710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2711_cast_fp16 = mul(x = var_2709_cast_fp16, y = var_2710_to_fp16)[name = string("op_2711_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2712_cast_fp16 = add(x = var_69_cast_fp16, y = var_2711_cast_fp16)[name = string("op_2712_cast_fp16")];
            fp32 var_2713_epsilon_0 = const()[name = string("op_2713_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2713_cast_fp16 = rsqrt(epsilon = var_2713_epsilon_0, x = var_2712_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_69_cast_fp16 = mul(x = xn_69_cast_fp16, y = var_2713_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390464704)))];
            tensor<fp16, [4, 64, 896]> hidden_states_239_cast_fp16 = mul(x = lm_layers_17_input_layernorm_weight_to_fp16, y = out_69_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390466560))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391269440))))[name = string("lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391271296)))];
            tensor<fp16, [4, 64, 896]> linear_119_cast_fp16 = linear(bias = lm_layers_17_self_attn_q_proj_bias_to_fp16, weight = lm_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_2725 = const()[name = string("op_2725"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2726_cast_fp16 = reshape(shape = var_2725, x = linear_119_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> q_35_perm_0 = const()[name = string("q_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391273152))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391387904))))[name = string("lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388224)))];
            tensor<fp16, [4, 64, 128]> linear_120_cast_fp16 = linear(bias = lm_layers_17_self_attn_k_proj_bias_to_fp16, weight = lm_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_2731 = const()[name = string("op_2731"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2732_cast_fp16 = reshape(shape = var_2731, x = linear_120_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391388544))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503296))))[name = string("lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503616)))];
            tensor<fp16, [4, 64, 128]> linear_121_cast_fp16 = linear(bias = lm_layers_17_self_attn_v_proj_bias_to_fp16, weight = lm_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_239_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_2737 = const()[name = string("op_2737"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2738_cast_fp16 = reshape(shape = var_2737, x = linear_121_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> hidden_states_245_perm_0 = const()[name = string("hidden_states_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_35_cast_fp16 = transpose(perm = q_35_perm_0, x = var_2726_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [4, 14, 64, 64]> var_2742_cast_fp16 = mul(x = q_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_35_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_35_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_321_promoted_to_fp16 = const()[name = string("const_321_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2753_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_321_promoted_to_fp16)[name = string("op_2753_cast_fp16")];
            bool var_2755_interleave_0 = const()[name = string("op_2755_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2755_cast_fp16 = concat(axis = var_20, interleave = var_2755_interleave_0, values = (var_2753_cast_fp16, x1_69_cast_fp16))[name = string("op_2755_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2756_cast_fp16 = mul(x = var_2755_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2756_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_35_cast_fp16 = add(x = var_2742_cast_fp16, y = var_2756_cast_fp16)[name = string("query_35_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = var_2732_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [4, 2, 64, 64]> var_2758_cast_fp16 = mul(x = k_35_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_35_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_35_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_324_promoted_to_fp16 = const()[name = string("const_324_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2769_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_324_promoted_to_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2771_interleave_0 = const()[name = string("op_2771_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2771_cast_fp16 = concat(axis = var_20, interleave = var_2771_interleave_0, values = (var_2769_cast_fp16, x1_71_cast_fp16))[name = string("op_2771_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2772_cast_fp16 = mul(x = var_2771_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2772_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_241_cast_fp16 = add(x = var_2758_cast_fp16, y = var_2772_cast_fp16)[name = string("hidden_states_241_cast_fp16")];
            tensor<int32, [1]> var_2781_axes_0 = const()[name = string("op_2781_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2781_cast_fp16 = expand_dims(axes = var_2781_axes_0, x = hidden_states_241_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<int32, [5]> hidden_states_243_reps_0 = const()[name = string("hidden_states_243_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_243_cast_fp16 = tile(reps = hidden_states_243_reps_0, x = var_2781_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<int32, [4]> var_2789 = const()[name = string("op_2789"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_35_cast_fp16 = reshape(shape = var_2789, x = hidden_states_243_cast_fp16)[name = string("key_35_cast_fp16")];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_245_cast_fp16 = transpose(perm = hidden_states_245_perm_0, x = var_2738_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2798_cast_fp16 = expand_dims(axes = var_2798_axes_0, x = hidden_states_245_cast_fp16)[name = string("op_2798_cast_fp16")];
            tensor<int32, [5]> hidden_states_247_reps_0 = const()[name = string("hidden_states_247_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_247_cast_fp16 = tile(reps = hidden_states_247_reps_0, x = var_2798_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            tensor<int32, [4]> var_2806 = const()[name = string("op_2806"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_35_cast_fp16 = reshape(shape = var_2806, x = hidden_states_247_cast_fp16)[name = string("value_35_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_35_cast_fp16 = mul(x = query_35_cast_fp16, y = var_13_to_fp16)[name = string("mul_35_cast_fp16")];
            bool matmul_17_transpose_y_0 = const()[name = string("matmul_17_transpose_y_0"), val = bool(true)];
            bool matmul_17_transpose_x_0 = const()[name = string("matmul_17_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_0, transpose_y = matmul_17_transpose_y_0, x = mul_35_cast_fp16, y = key_35_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_17_cast_fp16 = add(x = matmul_17_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_17_cast_fp16")];
            int32 softmax_17_axis_0 = const()[name = string("softmax_17_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_17_cast_fp16 = softmax(axis = softmax_17_axis_0, x = add_17_cast_fp16)[name = string("softmax_17_cast_fp16")];
            bool attn_output_69_transpose_x_0 = const()[name = string("attn_output_69_transpose_x_0"), val = bool(false)];
            bool attn_output_69_transpose_y_0 = const()[name = string("attn_output_69_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_69_cast_fp16 = matmul(transpose_x = attn_output_69_transpose_x_0, transpose_y = attn_output_69_transpose_y_0, x = softmax_17_cast_fp16, y = value_35_cast_fp16)[name = string("attn_output_69_cast_fp16")];
            tensor<int32, [4]> var_2809_perm_0 = const()[name = string("op_2809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2811 = const()[name = string("op_2811"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2809_cast_fp16 = transpose(perm = var_2809_perm_0, x = attn_output_69_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [4, 64, 896]> var_2812_cast_fp16 = reshape(shape = var_2811, x = var_2809_cast_fp16)[name = string("op_2812_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391503936))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392306816))))[name = string("lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_122_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = var_2812_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_69_cast_fp16 = add(x = x_67_cast_fp16, y = linear_122_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2818_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2818_cast_fp16")];
            tensor<int32, [1]> var_2820_axes_0 = const()[name = string("op_2820_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2820_cast_fp16 = reduce_max(axes = var_2820_axes_0, keep_dims = var_12, x = var_2818_cast_fp16)[name = string("op_2820_cast_fp16")];
            fp16 const_333_to_fp16 = const()[name = string("const_333_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_35_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_333_to_fp16, x = var_2820_cast_fp16)[name = string("clip_35_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = clip_35_cast_fp16)[name = string("xn_71_cast_fp16")];
            fp16 var_23_promoted_35_to_fp16 = const()[name = string("op_23_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2823_cast_fp16 = pow(x = xn_71_cast_fp16, y = var_23_promoted_35_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_71_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2826_cast_fp16 = mul(x = clip_35_cast_fp16, y = clip_35_cast_fp16)[name = string("op_2826_cast_fp16")];
            fp32 var_2827_epsilon_0 = const()[name = string("op_2827_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2827_cast_fp16 = inverse(epsilon = var_2827_epsilon_0, x = var_2826_cast_fp16)[name = string("op_2827_cast_fp16")];
            fp16 var_2828_to_fp16 = const()[name = string("op_2828_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2829_cast_fp16 = mul(x = var_2827_cast_fp16, y = var_2828_to_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2830_cast_fp16 = add(x = var_71_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2830_cast_fp16")];
            fp32 var_2831_epsilon_0 = const()[name = string("op_2831_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2831_cast_fp16 = rsqrt(epsilon = var_2831_epsilon_0, x = var_2830_cast_fp16)[name = string("op_2831_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_71_cast_fp16 = mul(x = xn_71_cast_fp16, y = var_2831_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [896]> lm_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392308672)))];
            tensor<fp16, [4, 64, 896]> input_139_cast_fp16 = mul(x = lm_layers_17_post_attention_layernorm_weight_to_fp16, y = out_71_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392310528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396668736))))[name = string("lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2839_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_2839_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396678528))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401036736))))[name = string("lm_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_143_cast_fp16 = mul(x = var_2839_cast_fp16, y = linear_124_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401046528))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405404736))))[name = string("lm_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_125_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_71_cast_fp16 = add(x = x_69_cast_fp16, y = linear_125_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2851_cast_fp16 = abs(x = x_71_cast_fp16)[name = string("op_2851_cast_fp16")];
            tensor<int32, [1]> var_2853_axes_0 = const()[name = string("op_2853_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2853_cast_fp16 = reduce_max(axes = var_2853_axes_0, keep_dims = var_12, x = var_2851_cast_fp16)[name = string("op_2853_cast_fp16")];
            fp16 const_334_to_fp16 = const()[name = string("const_334_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_36_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_334_to_fp16, x = var_2853_cast_fp16)[name = string("clip_36_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_73_cast_fp16 = real_div(x = x_71_cast_fp16, y = clip_36_cast_fp16)[name = string("xn_73_cast_fp16")];
            fp16 var_23_promoted_36_to_fp16 = const()[name = string("op_23_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2856_cast_fp16 = pow(x = xn_73_cast_fp16, y = var_23_promoted_36_to_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_2856_cast_fp16)[name = string("var_73_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2859_cast_fp16 = mul(x = clip_36_cast_fp16, y = clip_36_cast_fp16)[name = string("op_2859_cast_fp16")];
            fp32 var_2860_epsilon_0 = const()[name = string("op_2860_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2860_cast_fp16 = inverse(epsilon = var_2860_epsilon_0, x = var_2859_cast_fp16)[name = string("op_2860_cast_fp16")];
            fp16 var_2861_to_fp16 = const()[name = string("op_2861_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2862_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2861_to_fp16)[name = string("op_2862_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2863_cast_fp16 = add(x = var_73_cast_fp16, y = var_2862_cast_fp16)[name = string("op_2863_cast_fp16")];
            fp32 var_2864_epsilon_0 = const()[name = string("op_2864_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2864_cast_fp16 = rsqrt(epsilon = var_2864_epsilon_0, x = var_2863_cast_fp16)[name = string("op_2864_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_73_cast_fp16 = mul(x = xn_73_cast_fp16, y = var_2864_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405406592)))];
            tensor<fp16, [4, 64, 896]> hidden_states_253_cast_fp16 = mul(x = lm_layers_18_input_layernorm_weight_to_fp16, y = out_73_cast_fp16)[name = string("hidden_states_253_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405408448))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406211328))))[name = string("lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406213184)))];
            tensor<fp16, [4, 64, 896]> linear_126_cast_fp16 = linear(bias = lm_layers_18_self_attn_q_proj_bias_to_fp16, weight = lm_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_2876 = const()[name = string("op_2876"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_2877_cast_fp16 = reshape(shape = var_2876, x = linear_126_cast_fp16)[name = string("op_2877_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = string("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406215040))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406329792))))[name = string("lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330112)))];
            tensor<fp16, [4, 64, 128]> linear_127_cast_fp16 = linear(bias = lm_layers_18_self_attn_k_proj_bias_to_fp16, weight = lm_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_2882 = const()[name = string("op_2882"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2883_cast_fp16 = reshape(shape = var_2882, x = linear_127_cast_fp16)[name = string("op_2883_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = string("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406330432))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445184))))[name = string("lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445504)))];
            tensor<fp16, [4, 64, 128]> linear_128_cast_fp16 = linear(bias = lm_layers_18_self_attn_v_proj_bias_to_fp16, weight = lm_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_253_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_128_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> hidden_states_259_perm_0 = const()[name = string("hidden_states_259_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_2877_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [4, 14, 64, 64]> var_2893_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2893_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_37_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_37_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_339_promoted_to_fp16 = const()[name = string("const_339_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_2904_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_339_promoted_to_fp16)[name = string("op_2904_cast_fp16")];
            bool var_2906_interleave_0 = const()[name = string("op_2906_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_2906_cast_fp16 = concat(axis = var_20, interleave = var_2906_interleave_0, values = (var_2904_cast_fp16, x1_73_cast_fp16))[name = string("op_2906_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_2907_cast_fp16 = mul(x = var_2906_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_37_cast_fp16 = add(x = var_2893_cast_fp16, y = var_2907_cast_fp16)[name = string("query_37_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_2883_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [4, 2, 64, 64]> var_2909_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_5_cast_fp16)[name = string("op_2909_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_37_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_37_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_342_promoted_to_fp16 = const()[name = string("const_342_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_2920_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_342_promoted_to_fp16)[name = string("op_2920_cast_fp16")];
            bool var_2922_interleave_0 = const()[name = string("op_2922_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_2922_cast_fp16 = concat(axis = var_20, interleave = var_2922_interleave_0, values = (var_2920_cast_fp16, x1_75_cast_fp16))[name = string("op_2922_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_2923_cast_fp16 = mul(x = var_2922_cast_fp16, y = sin_5_cast_fp16)[name = string("op_2923_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_255_cast_fp16 = add(x = var_2909_cast_fp16, y = var_2923_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [1]> var_2932_axes_0 = const()[name = string("op_2932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2932_cast_fp16 = expand_dims(axes = var_2932_axes_0, x = hidden_states_255_cast_fp16)[name = string("op_2932_cast_fp16")];
            tensor<int32, [5]> hidden_states_257_reps_0 = const()[name = string("hidden_states_257_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_257_cast_fp16 = tile(reps = hidden_states_257_reps_0, x = var_2932_cast_fp16)[name = string("hidden_states_257_cast_fp16")];
            tensor<int32, [4]> var_2940 = const()[name = string("op_2940"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_37_cast_fp16 = reshape(shape = var_2940, x = hidden_states_257_cast_fp16)[name = string("key_37_cast_fp16")];
            tensor<int32, [1]> var_2949_axes_0 = const()[name = string("op_2949_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_259_cast_fp16 = transpose(perm = hidden_states_259_perm_0, x = var_2889_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_2949_cast_fp16 = expand_dims(axes = var_2949_axes_0, x = hidden_states_259_cast_fp16)[name = string("op_2949_cast_fp16")];
            tensor<int32, [5]> hidden_states_261_reps_0 = const()[name = string("hidden_states_261_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_261_cast_fp16 = tile(reps = hidden_states_261_reps_0, x = var_2949_cast_fp16)[name = string("hidden_states_261_cast_fp16")];
            tensor<int32, [4]> var_2957 = const()[name = string("op_2957"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_37_cast_fp16 = reshape(shape = var_2957, x = hidden_states_261_cast_fp16)[name = string("value_37_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_37_cast_fp16 = mul(x = query_37_cast_fp16, y = var_13_to_fp16)[name = string("mul_37_cast_fp16")];
            bool matmul_18_transpose_y_0 = const()[name = string("matmul_18_transpose_y_0"), val = bool(true)];
            bool matmul_18_transpose_x_0 = const()[name = string("matmul_18_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_0, transpose_y = matmul_18_transpose_y_0, x = mul_37_cast_fp16, y = key_37_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_18_cast_fp16 = add(x = matmul_18_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_18_cast_fp16")];
            int32 softmax_18_axis_0 = const()[name = string("softmax_18_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_18_cast_fp16 = softmax(axis = softmax_18_axis_0, x = add_18_cast_fp16)[name = string("softmax_18_cast_fp16")];
            bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)];
            bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_18_cast_fp16, y = value_37_cast_fp16)[name = string("attn_output_73_cast_fp16")];
            tensor<int32, [4]> var_2960_perm_0 = const()[name = string("op_2960_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2962 = const()[name = string("op_2962"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_2960_cast_fp16 = transpose(perm = var_2960_perm_0, x = attn_output_73_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [4, 64, 896]> var_2963_cast_fp16 = reshape(shape = var_2962, x = var_2960_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406445824))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407248704))))[name = string("lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_129_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = var_2963_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_129_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_2969_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2969_cast_fp16")];
            tensor<int32, [1]> var_2971_axes_0 = const()[name = string("op_2971_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_2971_cast_fp16 = reduce_max(axes = var_2971_axes_0, keep_dims = var_12, x = var_2969_cast_fp16)[name = string("op_2971_cast_fp16")];
            fp16 const_351_to_fp16 = const()[name = string("const_351_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_37_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_351_to_fp16, x = var_2971_cast_fp16)[name = string("clip_37_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_75_cast_fp16 = real_div(x = x_73_cast_fp16, y = clip_37_cast_fp16)[name = string("xn_75_cast_fp16")];
            fp16 var_23_promoted_37_to_fp16 = const()[name = string("op_23_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_2974_cast_fp16 = pow(x = xn_75_cast_fp16, y = var_23_promoted_37_to_fp16)[name = string("op_2974_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_2974_cast_fp16)[name = string("var_75_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2977_cast_fp16 = mul(x = clip_37_cast_fp16, y = clip_37_cast_fp16)[name = string("op_2977_cast_fp16")];
            fp32 var_2978_epsilon_0 = const()[name = string("op_2978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_2978_cast_fp16 = inverse(epsilon = var_2978_epsilon_0, x = var_2977_cast_fp16)[name = string("op_2978_cast_fp16")];
            fp16 var_2979_to_fp16 = const()[name = string("op_2979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_2980_cast_fp16 = mul(x = var_2978_cast_fp16, y = var_2979_to_fp16)[name = string("op_2980_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_2981_cast_fp16 = add(x = var_75_cast_fp16, y = var_2980_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp32 var_2982_epsilon_0 = const()[name = string("op_2982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_2982_cast_fp16 = rsqrt(epsilon = var_2982_epsilon_0, x = var_2981_cast_fp16)[name = string("op_2982_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_75_cast_fp16 = mul(x = xn_75_cast_fp16, y = var_2982_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [896]> lm_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407250560)))];
            tensor<fp16, [4, 64, 896]> input_147_cast_fp16 = mul(x = lm_layers_18_post_attention_layernorm_weight_to_fp16, y = out_75_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407252416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411610624))))[name = string("lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_2990_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_2990_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411620416))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415978624))))[name = string("lm_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_151_cast_fp16 = mul(x = var_2990_cast_fp16, y = linear_131_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415988416))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420346624))))[name = string("lm_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_132_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_75_cast_fp16 = add(x = x_73_cast_fp16, y = linear_132_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3002_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_3002_cast_fp16")];
            tensor<int32, [1]> var_3004_axes_0 = const()[name = string("op_3004_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3004_cast_fp16 = reduce_max(axes = var_3004_axes_0, keep_dims = var_12, x = var_3002_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_38_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_352_to_fp16, x = var_3004_cast_fp16)[name = string("clip_38_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_77_cast_fp16 = real_div(x = x_75_cast_fp16, y = clip_38_cast_fp16)[name = string("xn_77_cast_fp16")];
            fp16 var_23_promoted_38_to_fp16 = const()[name = string("op_23_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3007_cast_fp16 = pow(x = xn_77_cast_fp16, y = var_23_promoted_38_to_fp16)[name = string("op_3007_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_3007_cast_fp16)[name = string("var_77_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3010_cast_fp16 = mul(x = clip_38_cast_fp16, y = clip_38_cast_fp16)[name = string("op_3010_cast_fp16")];
            fp32 var_3011_epsilon_0 = const()[name = string("op_3011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3011_cast_fp16 = inverse(epsilon = var_3011_epsilon_0, x = var_3010_cast_fp16)[name = string("op_3011_cast_fp16")];
            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3013_cast_fp16 = mul(x = var_3011_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3014_cast_fp16 = add(x = var_77_cast_fp16, y = var_3013_cast_fp16)[name = string("op_3014_cast_fp16")];
            fp32 var_3015_epsilon_0 = const()[name = string("op_3015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3015_cast_fp16 = rsqrt(epsilon = var_3015_epsilon_0, x = var_3014_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_77_cast_fp16 = mul(x = xn_77_cast_fp16, y = var_3015_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420348480)))];
            tensor<fp16, [4, 64, 896]> hidden_states_267_cast_fp16 = mul(x = lm_layers_19_input_layernorm_weight_to_fp16, y = out_77_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420350336))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421153216))))[name = string("lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421155072)))];
            tensor<fp16, [4, 64, 896]> linear_133_cast_fp16 = linear(bias = lm_layers_19_self_attn_q_proj_bias_to_fp16, weight = lm_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3027 = const()[name = string("op_3027"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_3028_cast_fp16 = reshape(shape = var_3027, x = linear_133_cast_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421156928))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421271680))))[name = string("lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272000)))];
            tensor<fp16, [4, 64, 128]> linear_134_cast_fp16 = linear(bias = lm_layers_19_self_attn_k_proj_bias_to_fp16, weight = lm_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3034_cast_fp16 = reshape(shape = var_3033, x = linear_134_cast_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421272320))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387072))))[name = string("lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387392)))];
            tensor<fp16, [4, 64, 128]> linear_135_cast_fp16 = linear(bias = lm_layers_19_self_attn_v_proj_bias_to_fp16, weight = lm_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_267_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3039 = const()[name = string("op_3039"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3040_cast_fp16 = reshape(shape = var_3039, x = linear_135_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> hidden_states_273_perm_0 = const()[name = string("hidden_states_273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_3028_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [4, 14, 64, 64]> var_3044_cast_fp16 = mul(x = q_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_39_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_39_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_357_promoted_to_fp16 = const()[name = string("const_357_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_3055_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_357_promoted_to_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3057_interleave_0 = const()[name = string("op_3057_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_3057_cast_fp16 = concat(axis = var_20, interleave = var_3057_interleave_0, values = (var_3055_cast_fp16, x1_77_cast_fp16))[name = string("op_3057_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_3058_cast_fp16 = mul(x = var_3057_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_39_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3058_cast_fp16)[name = string("query_39_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_3034_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [4, 2, 64, 64]> var_3060_cast_fp16 = mul(x = k_39_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3060_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_39_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_39_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_360_promoted_to_fp16 = const()[name = string("const_360_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_3071_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_360_promoted_to_fp16)[name = string("op_3071_cast_fp16")];
            bool var_3073_interleave_0 = const()[name = string("op_3073_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_3073_cast_fp16 = concat(axis = var_20, interleave = var_3073_interleave_0, values = (var_3071_cast_fp16, x1_79_cast_fp16))[name = string("op_3073_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_3074_cast_fp16 = mul(x = var_3073_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_269_cast_fp16 = add(x = var_3060_cast_fp16, y = var_3074_cast_fp16)[name = string("hidden_states_269_cast_fp16")];
            tensor<int32, [1]> var_3083_axes_0 = const()[name = string("op_3083_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3083_cast_fp16 = expand_dims(axes = var_3083_axes_0, x = hidden_states_269_cast_fp16)[name = string("op_3083_cast_fp16")];
            tensor<int32, [5]> hidden_states_271_reps_0 = const()[name = string("hidden_states_271_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_271_cast_fp16 = tile(reps = hidden_states_271_reps_0, x = var_3083_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> var_3091 = const()[name = string("op_3091"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_39_cast_fp16 = reshape(shape = var_3091, x = hidden_states_271_cast_fp16)[name = string("key_39_cast_fp16")];
            tensor<int32, [1]> var_3100_axes_0 = const()[name = string("op_3100_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_273_cast_fp16 = transpose(perm = hidden_states_273_perm_0, x = var_3040_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3100_cast_fp16 = expand_dims(axes = var_3100_axes_0, x = hidden_states_273_cast_fp16)[name = string("op_3100_cast_fp16")];
            tensor<int32, [5]> hidden_states_275_reps_0 = const()[name = string("hidden_states_275_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_275_cast_fp16 = tile(reps = hidden_states_275_reps_0, x = var_3100_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_39_cast_fp16 = reshape(shape = var_3108, x = hidden_states_275_cast_fp16)[name = string("value_39_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_39_cast_fp16 = mul(x = query_39_cast_fp16, y = var_13_to_fp16)[name = string("mul_39_cast_fp16")];
            bool matmul_19_transpose_y_0 = const()[name = string("matmul_19_transpose_y_0"), val = bool(true)];
            bool matmul_19_transpose_x_0 = const()[name = string("matmul_19_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_0, transpose_y = matmul_19_transpose_y_0, x = mul_39_cast_fp16, y = key_39_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_19_cast_fp16 = add(x = matmul_19_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_19_cast_fp16")];
            int32 softmax_19_axis_0 = const()[name = string("softmax_19_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_19_cast_fp16 = softmax(axis = softmax_19_axis_0, x = add_19_cast_fp16)[name = string("softmax_19_cast_fp16")];
            bool attn_output_77_transpose_x_0 = const()[name = string("attn_output_77_transpose_x_0"), val = bool(false)];
            bool attn_output_77_transpose_y_0 = const()[name = string("attn_output_77_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_77_cast_fp16 = matmul(transpose_x = attn_output_77_transpose_x_0, transpose_y = attn_output_77_transpose_y_0, x = softmax_19_cast_fp16, y = value_39_cast_fp16)[name = string("attn_output_77_cast_fp16")];
            tensor<int32, [4]> var_3111_perm_0 = const()[name = string("op_3111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3113 = const()[name = string("op_3113"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_3111_cast_fp16 = transpose(perm = var_3111_perm_0, x = attn_output_77_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [4, 64, 896]> var_3114_cast_fp16 = reshape(shape = var_3113, x = var_3111_cast_fp16)[name = string("op_3114_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421387712))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422190592))))[name = string("lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_136_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = var_3114_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_77_cast_fp16 = add(x = x_75_cast_fp16, y = linear_136_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3120_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<int32, [1]> var_3122_axes_0 = const()[name = string("op_3122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3122_cast_fp16 = reduce_max(axes = var_3122_axes_0, keep_dims = var_12, x = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            fp16 const_369_to_fp16 = const()[name = string("const_369_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_39_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_369_to_fp16, x = var_3122_cast_fp16)[name = string("clip_39_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = clip_39_cast_fp16)[name = string("xn_79_cast_fp16")];
            fp16 var_23_promoted_39_to_fp16 = const()[name = string("op_23_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3125_cast_fp16 = pow(x = xn_79_cast_fp16, y = var_23_promoted_39_to_fp16)[name = string("op_3125_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_3125_cast_fp16)[name = string("var_79_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3128_cast_fp16 = mul(x = clip_39_cast_fp16, y = clip_39_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3129_cast_fp16 = inverse(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3131_cast_fp16 = mul(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("op_3131_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3132_cast_fp16 = add(x = var_79_cast_fp16, y = var_3131_cast_fp16)[name = string("op_3132_cast_fp16")];
            fp32 var_3133_epsilon_0 = const()[name = string("op_3133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3133_cast_fp16 = rsqrt(epsilon = var_3133_epsilon_0, x = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_79_cast_fp16 = mul(x = xn_79_cast_fp16, y = var_3133_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [896]> lm_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422192448)))];
            tensor<fp16, [4, 64, 896]> input_155_cast_fp16 = mul(x = lm_layers_19_post_attention_layernorm_weight_to_fp16, y = out_79_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422194304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426552512))))[name = string("lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_3141_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_3141_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426562304))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430920512))))[name = string("lm_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_159_cast_fp16 = mul(x = var_3141_cast_fp16, y = linear_138_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430930304))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435288512))))[name = string("lm_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_139_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_79_cast_fp16 = add(x = x_77_cast_fp16, y = linear_139_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3153_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_3153_cast_fp16")];
            tensor<int32, [1]> var_3155_axes_0 = const()[name = string("op_3155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3155_cast_fp16 = reduce_max(axes = var_3155_axes_0, keep_dims = var_12, x = var_3153_cast_fp16)[name = string("op_3155_cast_fp16")];
            fp16 const_370_to_fp16 = const()[name = string("const_370_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_40_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_370_to_fp16, x = var_3155_cast_fp16)[name = string("clip_40_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_81_cast_fp16 = real_div(x = x_79_cast_fp16, y = clip_40_cast_fp16)[name = string("xn_81_cast_fp16")];
            fp16 var_23_promoted_40_to_fp16 = const()[name = string("op_23_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3158_cast_fp16 = pow(x = xn_81_cast_fp16, y = var_23_promoted_40_to_fp16)[name = string("op_3158_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_3158_cast_fp16)[name = string("var_81_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3161_cast_fp16 = mul(x = clip_40_cast_fp16, y = clip_40_cast_fp16)[name = string("op_3161_cast_fp16")];
            fp32 var_3162_epsilon_0 = const()[name = string("op_3162_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3162_cast_fp16 = inverse(epsilon = var_3162_epsilon_0, x = var_3161_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3163_to_fp16 = const()[name = string("op_3163_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3164_cast_fp16 = mul(x = var_3162_cast_fp16, y = var_3163_to_fp16)[name = string("op_3164_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3165_cast_fp16 = add(x = var_81_cast_fp16, y = var_3164_cast_fp16)[name = string("op_3165_cast_fp16")];
            fp32 var_3166_epsilon_0 = const()[name = string("op_3166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3166_cast_fp16 = rsqrt(epsilon = var_3166_epsilon_0, x = var_3165_cast_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_81_cast_fp16 = mul(x = xn_81_cast_fp16, y = var_3166_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435290368)))];
            tensor<fp16, [4, 64, 896]> hidden_states_281_cast_fp16 = mul(x = lm_layers_20_input_layernorm_weight_to_fp16, y = out_81_cast_fp16)[name = string("hidden_states_281_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435292224))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436095104))))[name = string("lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436096960)))];
            tensor<fp16, [4, 64, 896]> linear_140_cast_fp16 = linear(bias = lm_layers_20_self_attn_q_proj_bias_to_fp16, weight = lm_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_140_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = string("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436098816))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213568))))[name = string("lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436213888)))];
            tensor<fp16, [4, 64, 128]> linear_141_cast_fp16 = linear(bias = lm_layers_20_self_attn_k_proj_bias_to_fp16, weight = lm_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3184 = const()[name = string("op_3184"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3185_cast_fp16 = reshape(shape = var_3184, x = linear_141_cast_fp16)[name = string("op_3185_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = string("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436214208))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436328960))))[name = string("lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329280)))];
            tensor<fp16, [4, 64, 128]> linear_142_cast_fp16 = linear(bias = lm_layers_20_self_attn_v_proj_bias_to_fp16, weight = lm_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_281_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3190 = const()[name = string("op_3190"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3191_cast_fp16 = reshape(shape = var_3190, x = linear_142_cast_fp16)[name = string("op_3191_cast_fp16")];
            tensor<int32, [4]> hidden_states_287_perm_0 = const()[name = string("hidden_states_287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_3179_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [4, 14, 64, 64]> var_3195_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3195_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_41_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_41_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_375_promoted_to_fp16 = const()[name = string("const_375_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_3206_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_375_promoted_to_fp16)[name = string("op_3206_cast_fp16")];
            bool var_3208_interleave_0 = const()[name = string("op_3208_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_3208_cast_fp16 = concat(axis = var_20, interleave = var_3208_interleave_0, values = (var_3206_cast_fp16, x1_81_cast_fp16))[name = string("op_3208_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_3209_cast_fp16 = mul(x = var_3208_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3209_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_41_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3209_cast_fp16)[name = string("query_41_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_3185_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [4, 2, 64, 64]> var_3211_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3211_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_41_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_41_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_378_promoted_to_fp16 = const()[name = string("const_378_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_3222_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_378_promoted_to_fp16)[name = string("op_3222_cast_fp16")];
            bool var_3224_interleave_0 = const()[name = string("op_3224_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_3224_cast_fp16 = concat(axis = var_20, interleave = var_3224_interleave_0, values = (var_3222_cast_fp16, x1_83_cast_fp16))[name = string("op_3224_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_3225_cast_fp16 = mul(x = var_3224_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_283_cast_fp16 = add(x = var_3211_cast_fp16, y = var_3225_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<int32, [1]> var_3234_axes_0 = const()[name = string("op_3234_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3234_cast_fp16 = expand_dims(axes = var_3234_axes_0, x = hidden_states_283_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<int32, [5]> hidden_states_285_reps_0 = const()[name = string("hidden_states_285_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_285_cast_fp16 = tile(reps = hidden_states_285_reps_0, x = var_3234_cast_fp16)[name = string("hidden_states_285_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = string("op_3242"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_41_cast_fp16 = reshape(shape = var_3242, x = hidden_states_285_cast_fp16)[name = string("key_41_cast_fp16")];
            tensor<int32, [1]> var_3251_axes_0 = const()[name = string("op_3251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_287_cast_fp16 = transpose(perm = hidden_states_287_perm_0, x = var_3191_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3251_cast_fp16 = expand_dims(axes = var_3251_axes_0, x = hidden_states_287_cast_fp16)[name = string("op_3251_cast_fp16")];
            tensor<int32, [5]> hidden_states_289_reps_0 = const()[name = string("hidden_states_289_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_289_cast_fp16 = tile(reps = hidden_states_289_reps_0, x = var_3251_cast_fp16)[name = string("hidden_states_289_cast_fp16")];
            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_41_cast_fp16 = reshape(shape = var_3259, x = hidden_states_289_cast_fp16)[name = string("value_41_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_41_cast_fp16 = mul(x = query_41_cast_fp16, y = var_13_to_fp16)[name = string("mul_41_cast_fp16")];
            bool matmul_20_transpose_y_0 = const()[name = string("matmul_20_transpose_y_0"), val = bool(true)];
            bool matmul_20_transpose_x_0 = const()[name = string("matmul_20_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_0, transpose_y = matmul_20_transpose_y_0, x = mul_41_cast_fp16, y = key_41_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_20_cast_fp16 = add(x = matmul_20_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_20_cast_fp16")];
            int32 softmax_20_axis_0 = const()[name = string("softmax_20_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_20_cast_fp16 = softmax(axis = softmax_20_axis_0, x = add_20_cast_fp16)[name = string("softmax_20_cast_fp16")];
            bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)];
            bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_20_cast_fp16, y = value_41_cast_fp16)[name = string("attn_output_81_cast_fp16")];
            tensor<int32, [4]> var_3262_perm_0 = const()[name = string("op_3262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_3262_cast_fp16 = transpose(perm = var_3262_perm_0, x = attn_output_81_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [4, 64, 896]> var_3265_cast_fp16 = reshape(shape = var_3264, x = var_3262_cast_fp16)[name = string("op_3265_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436329600))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437132480))))[name = string("lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_143_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = var_3265_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_81_cast_fp16 = add(x = x_79_cast_fp16, y = linear_143_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3271_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_3271_cast_fp16")];
            tensor<int32, [1]> var_3273_axes_0 = const()[name = string("op_3273_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3273_cast_fp16 = reduce_max(axes = var_3273_axes_0, keep_dims = var_12, x = var_3271_cast_fp16)[name = string("op_3273_cast_fp16")];
            fp16 const_387_to_fp16 = const()[name = string("const_387_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_41_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_387_to_fp16, x = var_3273_cast_fp16)[name = string("clip_41_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = clip_41_cast_fp16)[name = string("xn_83_cast_fp16")];
            fp16 var_23_promoted_41_to_fp16 = const()[name = string("op_23_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3276_cast_fp16 = pow(x = xn_83_cast_fp16, y = var_23_promoted_41_to_fp16)[name = string("op_3276_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_83_cast_fp16 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_3276_cast_fp16)[name = string("var_83_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3279_cast_fp16 = mul(x = clip_41_cast_fp16, y = clip_41_cast_fp16)[name = string("op_3279_cast_fp16")];
            fp32 var_3280_epsilon_0 = const()[name = string("op_3280_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3280_cast_fp16 = inverse(epsilon = var_3280_epsilon_0, x = var_3279_cast_fp16)[name = string("op_3280_cast_fp16")];
            fp16 var_3281_to_fp16 = const()[name = string("op_3281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3282_cast_fp16 = mul(x = var_3280_cast_fp16, y = var_3281_to_fp16)[name = string("op_3282_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3283_cast_fp16 = add(x = var_83_cast_fp16, y = var_3282_cast_fp16)[name = string("op_3283_cast_fp16")];
            fp32 var_3284_epsilon_0 = const()[name = string("op_3284_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3284_cast_fp16 = rsqrt(epsilon = var_3284_epsilon_0, x = var_3283_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_83_cast_fp16 = mul(x = xn_83_cast_fp16, y = var_3284_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [896]> lm_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437134336)))];
            tensor<fp16, [4, 64, 896]> input_163_cast_fp16 = mul(x = lm_layers_20_post_attention_layernorm_weight_to_fp16, y = out_83_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437136192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441494400))))[name = string("lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_3292_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_3292_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441504192))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445862400))))[name = string("lm_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_167_cast_fp16 = mul(x = var_3292_cast_fp16, y = linear_145_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445872192))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450230400))))[name = string("lm_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_146_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_83_cast_fp16 = add(x = x_81_cast_fp16, y = linear_146_cast_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3304_cast_fp16 = abs(x = x_83_cast_fp16)[name = string("op_3304_cast_fp16")];
            tensor<int32, [1]> var_3306_axes_0 = const()[name = string("op_3306_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3306_cast_fp16 = reduce_max(axes = var_3306_axes_0, keep_dims = var_12, x = var_3304_cast_fp16)[name = string("op_3306_cast_fp16")];
            fp16 const_388_to_fp16 = const()[name = string("const_388_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_42_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_388_to_fp16, x = var_3306_cast_fp16)[name = string("clip_42_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_85_cast_fp16 = real_div(x = x_83_cast_fp16, y = clip_42_cast_fp16)[name = string("xn_85_cast_fp16")];
            fp16 var_23_promoted_42_to_fp16 = const()[name = string("op_23_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3309_cast_fp16 = pow(x = xn_85_cast_fp16, y = var_23_promoted_42_to_fp16)[name = string("op_3309_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_3309_cast_fp16)[name = string("var_85_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3312_cast_fp16 = mul(x = clip_42_cast_fp16, y = clip_42_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp32 var_3313_epsilon_0 = const()[name = string("op_3313_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3313_cast_fp16 = inverse(epsilon = var_3313_epsilon_0, x = var_3312_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3315_cast_fp16 = mul(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3316_cast_fp16 = add(x = var_85_cast_fp16, y = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            fp32 var_3317_epsilon_0 = const()[name = string("op_3317_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3317_cast_fp16 = rsqrt(epsilon = var_3317_epsilon_0, x = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_85_cast_fp16 = mul(x = xn_85_cast_fp16, y = var_3317_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232256)))];
            tensor<fp16, [4, 64, 896]> hidden_states_295_cast_fp16 = mul(x = lm_layers_21_input_layernorm_weight_to_fp16, y = out_85_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451036992))))[name = string("lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451038848)))];
            tensor<fp16, [4, 64, 896]> linear_147_cast_fp16 = linear(bias = lm_layers_21_self_attn_q_proj_bias_to_fp16, weight = lm_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3329 = const()[name = string("op_3329"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_3330_cast_fp16 = reshape(shape = var_3329, x = linear_147_cast_fp16)[name = string("op_3330_cast_fp16")];
            tensor<int32, [4]> q_43_perm_0 = const()[name = string("q_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451040704))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155456))))[name = string("lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451155776)))];
            tensor<fp16, [4, 64, 128]> linear_148_cast_fp16 = linear(bias = lm_layers_21_self_attn_k_proj_bias_to_fp16, weight = lm_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3335 = const()[name = string("op_3335"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3336_cast_fp16 = reshape(shape = var_3335, x = linear_148_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451156096))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451270848))))[name = string("lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271168)))];
            tensor<fp16, [4, 64, 128]> linear_149_cast_fp16 = linear(bias = lm_layers_21_self_attn_v_proj_bias_to_fp16, weight = lm_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_295_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3341 = const()[name = string("op_3341"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3342_cast_fp16 = reshape(shape = var_3341, x = linear_149_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<int32, [4]> hidden_states_301_perm_0 = const()[name = string("hidden_states_301_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_43_cast_fp16 = transpose(perm = q_43_perm_0, x = var_3330_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [4, 14, 64, 64]> var_3346_cast_fp16 = mul(x = q_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_43_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_43_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_393_promoted_to_fp16 = const()[name = string("const_393_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_3357_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_393_promoted_to_fp16)[name = string("op_3357_cast_fp16")];
            bool var_3359_interleave_0 = const()[name = string("op_3359_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_3359_cast_fp16 = concat(axis = var_20, interleave = var_3359_interleave_0, values = (var_3357_cast_fp16, x1_85_cast_fp16))[name = string("op_3359_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_3360_cast_fp16 = mul(x = var_3359_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3360_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_43_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3360_cast_fp16)[name = string("query_43_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = var_3336_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [4, 2, 64, 64]> var_3362_cast_fp16 = mul(x = k_43_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3362_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_43_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_43_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_396_promoted_to_fp16 = const()[name = string("const_396_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_3373_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_396_promoted_to_fp16)[name = string("op_3373_cast_fp16")];
            bool var_3375_interleave_0 = const()[name = string("op_3375_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_3375_cast_fp16 = concat(axis = var_20, interleave = var_3375_interleave_0, values = (var_3373_cast_fp16, x1_87_cast_fp16))[name = string("op_3375_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_3376_cast_fp16 = mul(x = var_3375_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3376_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_297_cast_fp16 = add(x = var_3362_cast_fp16, y = var_3376_cast_fp16)[name = string("hidden_states_297_cast_fp16")];
            tensor<int32, [1]> var_3385_axes_0 = const()[name = string("op_3385_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3385_cast_fp16 = expand_dims(axes = var_3385_axes_0, x = hidden_states_297_cast_fp16)[name = string("op_3385_cast_fp16")];
            tensor<int32, [5]> hidden_states_299_reps_0 = const()[name = string("hidden_states_299_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_299_cast_fp16 = tile(reps = hidden_states_299_reps_0, x = var_3385_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_43_cast_fp16 = reshape(shape = var_3393, x = hidden_states_299_cast_fp16)[name = string("key_43_cast_fp16")];
            tensor<int32, [1]> var_3402_axes_0 = const()[name = string("op_3402_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_301_cast_fp16 = transpose(perm = hidden_states_301_perm_0, x = var_3342_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3402_cast_fp16 = expand_dims(axes = var_3402_axes_0, x = hidden_states_301_cast_fp16)[name = string("op_3402_cast_fp16")];
            tensor<int32, [5]> hidden_states_303_reps_0 = const()[name = string("hidden_states_303_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_303_cast_fp16 = tile(reps = hidden_states_303_reps_0, x = var_3402_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> var_3410 = const()[name = string("op_3410"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_43_cast_fp16 = reshape(shape = var_3410, x = hidden_states_303_cast_fp16)[name = string("value_43_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_43_cast_fp16 = mul(x = query_43_cast_fp16, y = var_13_to_fp16)[name = string("mul_43_cast_fp16")];
            bool matmul_21_transpose_y_0 = const()[name = string("matmul_21_transpose_y_0"), val = bool(true)];
            bool matmul_21_transpose_x_0 = const()[name = string("matmul_21_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_0, transpose_y = matmul_21_transpose_y_0, x = mul_43_cast_fp16, y = key_43_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_21_cast_fp16 = add(x = matmul_21_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_21_cast_fp16")];
            int32 softmax_21_axis_0 = const()[name = string("softmax_21_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_21_cast_fp16 = softmax(axis = softmax_21_axis_0, x = add_21_cast_fp16)[name = string("softmax_21_cast_fp16")];
            bool attn_output_85_transpose_x_0 = const()[name = string("attn_output_85_transpose_x_0"), val = bool(false)];
            bool attn_output_85_transpose_y_0 = const()[name = string("attn_output_85_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_85_cast_fp16 = matmul(transpose_x = attn_output_85_transpose_x_0, transpose_y = attn_output_85_transpose_y_0, x = softmax_21_cast_fp16, y = value_43_cast_fp16)[name = string("attn_output_85_cast_fp16")];
            tensor<int32, [4]> var_3413_perm_0 = const()[name = string("op_3413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3415 = const()[name = string("op_3415"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_3413_cast_fp16 = transpose(perm = var_3413_perm_0, x = attn_output_85_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [4, 64, 896]> var_3416_cast_fp16 = reshape(shape = var_3415, x = var_3413_cast_fp16)[name = string("op_3416_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(451271488))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452074368))))[name = string("lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_150_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = var_3416_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_150_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3422_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [1]> var_3424_axes_0 = const()[name = string("op_3424_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3424_cast_fp16 = reduce_max(axes = var_3424_axes_0, keep_dims = var_12, x = var_3422_cast_fp16)[name = string("op_3424_cast_fp16")];
            fp16 const_405_to_fp16 = const()[name = string("const_405_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_43_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_405_to_fp16, x = var_3424_cast_fp16)[name = string("clip_43_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = clip_43_cast_fp16)[name = string("xn_87_cast_fp16")];
            fp16 var_23_promoted_43_to_fp16 = const()[name = string("op_23_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3427_cast_fp16 = pow(x = xn_87_cast_fp16, y = var_23_promoted_43_to_fp16)[name = string("op_3427_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_3427_cast_fp16)[name = string("var_87_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3430_cast_fp16 = mul(x = clip_43_cast_fp16, y = clip_43_cast_fp16)[name = string("op_3430_cast_fp16")];
            fp32 var_3431_epsilon_0 = const()[name = string("op_3431_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3431_cast_fp16 = inverse(epsilon = var_3431_epsilon_0, x = var_3430_cast_fp16)[name = string("op_3431_cast_fp16")];
            fp16 var_3432_to_fp16 = const()[name = string("op_3432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3433_cast_fp16 = mul(x = var_3431_cast_fp16, y = var_3432_to_fp16)[name = string("op_3433_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3434_cast_fp16 = add(x = var_87_cast_fp16, y = var_3433_cast_fp16)[name = string("op_3434_cast_fp16")];
            fp32 var_3435_epsilon_0 = const()[name = string("op_3435_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3435_cast_fp16 = rsqrt(epsilon = var_3435_epsilon_0, x = var_3434_cast_fp16)[name = string("op_3435_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_87_cast_fp16 = mul(x = xn_87_cast_fp16, y = var_3435_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [896]> lm_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452076224)))];
            tensor<fp16, [4, 64, 896]> input_171_cast_fp16 = mul(x = lm_layers_21_post_attention_layernorm_weight_to_fp16, y = out_87_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(452078080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456436288))))[name = string("lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_3443_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_3443_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456446080))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460804288))))[name = string("lm_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_175_cast_fp16 = mul(x = var_3443_cast_fp16, y = linear_152_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460814080))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465172288))))[name = string("lm_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_153_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_87_cast_fp16 = add(x = x_85_cast_fp16, y = linear_153_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3455_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3455_cast_fp16")];
            tensor<int32, [1]> var_3457_axes_0 = const()[name = string("op_3457_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3457_cast_fp16 = reduce_max(axes = var_3457_axes_0, keep_dims = var_12, x = var_3455_cast_fp16)[name = string("op_3457_cast_fp16")];
            fp16 const_406_to_fp16 = const()[name = string("const_406_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_44_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_406_to_fp16, x = var_3457_cast_fp16)[name = string("clip_44_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_89_cast_fp16 = real_div(x = x_87_cast_fp16, y = clip_44_cast_fp16)[name = string("xn_89_cast_fp16")];
            fp16 var_23_promoted_44_to_fp16 = const()[name = string("op_23_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3460_cast_fp16 = pow(x = xn_89_cast_fp16, y = var_23_promoted_44_to_fp16)[name = string("op_3460_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_3460_cast_fp16)[name = string("var_89_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3463_cast_fp16 = mul(x = clip_44_cast_fp16, y = clip_44_cast_fp16)[name = string("op_3463_cast_fp16")];
            fp32 var_3464_epsilon_0 = const()[name = string("op_3464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3464_cast_fp16 = inverse(epsilon = var_3464_epsilon_0, x = var_3463_cast_fp16)[name = string("op_3464_cast_fp16")];
            fp16 var_3465_to_fp16 = const()[name = string("op_3465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3466_cast_fp16 = mul(x = var_3464_cast_fp16, y = var_3465_to_fp16)[name = string("op_3466_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3467_cast_fp16 = add(x = var_89_cast_fp16, y = var_3466_cast_fp16)[name = string("op_3467_cast_fp16")];
            fp32 var_3468_epsilon_0 = const()[name = string("op_3468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3468_cast_fp16 = rsqrt(epsilon = var_3468_epsilon_0, x = var_3467_cast_fp16)[name = string("op_3468_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_89_cast_fp16 = mul(x = xn_89_cast_fp16, y = var_3468_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465174144)))];
            tensor<fp16, [4, 64, 896]> hidden_states_309_cast_fp16 = mul(x = lm_layers_22_input_layernorm_weight_to_fp16, y = out_89_cast_fp16)[name = string("hidden_states_309_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465176000))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465978880))))[name = string("lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465980736)))];
            tensor<fp16, [4, 64, 896]> linear_154_cast_fp16 = linear(bias = lm_layers_22_self_attn_q_proj_bias_to_fp16, weight = lm_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_3481_cast_fp16 = reshape(shape = var_3480, x = linear_154_cast_fp16)[name = string("op_3481_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465982592))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097344))))[name = string("lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097664)))];
            tensor<fp16, [4, 64, 128]> linear_155_cast_fp16 = linear(bias = lm_layers_22_self_attn_k_proj_bias_to_fp16, weight = lm_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3486 = const()[name = string("op_3486"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3487_cast_fp16 = reshape(shape = var_3486, x = linear_155_cast_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = string("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466097984))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466212736))))[name = string("lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213056)))];
            tensor<fp16, [4, 64, 128]> linear_156_cast_fp16 = linear(bias = lm_layers_22_self_attn_v_proj_bias_to_fp16, weight = lm_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_309_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3492 = const()[name = string("op_3492"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3493_cast_fp16 = reshape(shape = var_3492, x = linear_156_cast_fp16)[name = string("op_3493_cast_fp16")];
            tensor<int32, [4]> hidden_states_315_perm_0 = const()[name = string("hidden_states_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_3481_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [4, 14, 64, 64]> var_3497_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_45_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_45_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_411_promoted_to_fp16 = const()[name = string("const_411_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_3508_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_411_promoted_to_fp16)[name = string("op_3508_cast_fp16")];
            bool var_3510_interleave_0 = const()[name = string("op_3510_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_3510_cast_fp16 = concat(axis = var_20, interleave = var_3510_interleave_0, values = (var_3508_cast_fp16, x1_89_cast_fp16))[name = string("op_3510_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_3511_cast_fp16 = mul(x = var_3510_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_45_cast_fp16 = add(x = var_3497_cast_fp16, y = var_3511_cast_fp16)[name = string("query_45_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_3487_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [4, 2, 64, 64]> var_3513_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3513_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_45_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_45_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_414_promoted_to_fp16 = const()[name = string("const_414_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_3524_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_414_promoted_to_fp16)[name = string("op_3524_cast_fp16")];
            bool var_3526_interleave_0 = const()[name = string("op_3526_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_3526_cast_fp16 = concat(axis = var_20, interleave = var_3526_interleave_0, values = (var_3524_cast_fp16, x1_91_cast_fp16))[name = string("op_3526_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_3527_cast_fp16 = mul(x = var_3526_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3527_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_311_cast_fp16 = add(x = var_3513_cast_fp16, y = var_3527_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            tensor<int32, [1]> var_3536_axes_0 = const()[name = string("op_3536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3536_cast_fp16 = expand_dims(axes = var_3536_axes_0, x = hidden_states_311_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<int32, [5]> hidden_states_313_reps_0 = const()[name = string("hidden_states_313_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_313_cast_fp16 = tile(reps = hidden_states_313_reps_0, x = var_3536_cast_fp16)[name = string("hidden_states_313_cast_fp16")];
            tensor<int32, [4]> var_3544 = const()[name = string("op_3544"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_45_cast_fp16 = reshape(shape = var_3544, x = hidden_states_313_cast_fp16)[name = string("key_45_cast_fp16")];
            tensor<int32, [1]> var_3553_axes_0 = const()[name = string("op_3553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_315_cast_fp16 = transpose(perm = hidden_states_315_perm_0, x = var_3493_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3553_cast_fp16 = expand_dims(axes = var_3553_axes_0, x = hidden_states_315_cast_fp16)[name = string("op_3553_cast_fp16")];
            tensor<int32, [5]> hidden_states_317_reps_0 = const()[name = string("hidden_states_317_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_317_cast_fp16 = tile(reps = hidden_states_317_reps_0, x = var_3553_cast_fp16)[name = string("hidden_states_317_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = string("op_3561"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_45_cast_fp16 = reshape(shape = var_3561, x = hidden_states_317_cast_fp16)[name = string("value_45_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_45_cast_fp16 = mul(x = query_45_cast_fp16, y = var_13_to_fp16)[name = string("mul_45_cast_fp16")];
            bool matmul_22_transpose_y_0 = const()[name = string("matmul_22_transpose_y_0"), val = bool(true)];
            bool matmul_22_transpose_x_0 = const()[name = string("matmul_22_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_0, transpose_y = matmul_22_transpose_y_0, x = mul_45_cast_fp16, y = key_45_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_22_cast_fp16 = add(x = matmul_22_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_22_cast_fp16")];
            int32 softmax_22_axis_0 = const()[name = string("softmax_22_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_22_cast_fp16 = softmax(axis = softmax_22_axis_0, x = add_22_cast_fp16)[name = string("softmax_22_cast_fp16")];
            bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)];
            bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_22_cast_fp16, y = value_45_cast_fp16)[name = string("attn_output_89_cast_fp16")];
            tensor<int32, [4]> var_3564_perm_0 = const()[name = string("op_3564_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_3564_cast_fp16 = transpose(perm = var_3564_perm_0, x = attn_output_89_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [4, 64, 896]> var_3567_cast_fp16 = reshape(shape = var_3566, x = var_3564_cast_fp16)[name = string("op_3567_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(466213376))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467016256))))[name = string("lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_157_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = var_3567_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_89_cast_fp16 = add(x = x_87_cast_fp16, y = linear_157_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3573_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_3573_cast_fp16")];
            tensor<int32, [1]> var_3575_axes_0 = const()[name = string("op_3575_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3575_cast_fp16 = reduce_max(axes = var_3575_axes_0, keep_dims = var_12, x = var_3573_cast_fp16)[name = string("op_3575_cast_fp16")];
            fp16 const_423_to_fp16 = const()[name = string("const_423_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_45_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_423_to_fp16, x = var_3575_cast_fp16)[name = string("clip_45_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = clip_45_cast_fp16)[name = string("xn_91_cast_fp16")];
            fp16 var_23_promoted_45_to_fp16 = const()[name = string("op_23_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3578_cast_fp16 = pow(x = xn_91_cast_fp16, y = var_23_promoted_45_to_fp16)[name = string("op_3578_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_3578_cast_fp16)[name = string("var_91_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3581_cast_fp16 = mul(x = clip_45_cast_fp16, y = clip_45_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp32 var_3582_epsilon_0 = const()[name = string("op_3582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3582_cast_fp16 = inverse(epsilon = var_3582_epsilon_0, x = var_3581_cast_fp16)[name = string("op_3582_cast_fp16")];
            fp16 var_3583_to_fp16 = const()[name = string("op_3583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3584_cast_fp16 = mul(x = var_3582_cast_fp16, y = var_3583_to_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3585_cast_fp16 = add(x = var_91_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            fp32 var_3586_epsilon_0 = const()[name = string("op_3586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3586_cast_fp16 = rsqrt(epsilon = var_3586_epsilon_0, x = var_3585_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_91_cast_fp16 = mul(x = xn_91_cast_fp16, y = var_3586_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [896]> lm_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467018112)))];
            tensor<fp16, [4, 64, 896]> input_179_cast_fp16 = mul(x = lm_layers_22_post_attention_layernorm_weight_to_fp16, y = out_91_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467019968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471378176))))[name = string("lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_3594_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_3594_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471387968))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475746176))))[name = string("lm_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_183_cast_fp16 = mul(x = var_3594_cast_fp16, y = linear_159_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475755968))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480114176))))[name = string("lm_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_160_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_91_cast_fp16 = add(x = x_89_cast_fp16, y = linear_160_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3606_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3608_cast_fp16 = reduce_max(axes = var_3608_axes_0, keep_dims = var_12, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 const_424_to_fp16 = const()[name = string("const_424_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_46_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_424_to_fp16, x = var_3608_cast_fp16)[name = string("clip_46_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_93_cast_fp16 = real_div(x = x_91_cast_fp16, y = clip_46_cast_fp16)[name = string("xn_93_cast_fp16")];
            fp16 var_23_promoted_46_to_fp16 = const()[name = string("op_23_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3611_cast_fp16 = pow(x = xn_93_cast_fp16, y = var_23_promoted_46_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_93_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3614_cast_fp16 = mul(x = clip_46_cast_fp16, y = clip_46_cast_fp16)[name = string("op_3614_cast_fp16")];
            fp32 var_3615_epsilon_0 = const()[name = string("op_3615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3615_cast_fp16 = inverse(epsilon = var_3615_epsilon_0, x = var_3614_cast_fp16)[name = string("op_3615_cast_fp16")];
            fp16 var_3616_to_fp16 = const()[name = string("op_3616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3617_cast_fp16 = mul(x = var_3615_cast_fp16, y = var_3616_to_fp16)[name = string("op_3617_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3618_cast_fp16 = add(x = var_93_cast_fp16, y = var_3617_cast_fp16)[name = string("op_3618_cast_fp16")];
            fp32 var_3619_epsilon_0 = const()[name = string("op_3619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3619_cast_fp16 = rsqrt(epsilon = var_3619_epsilon_0, x = var_3618_cast_fp16)[name = string("op_3619_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_93_cast_fp16 = mul(x = xn_93_cast_fp16, y = var_3619_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480116032)))];
            tensor<fp16, [4, 64, 896]> hidden_states_323_cast_fp16 = mul(x = lm_layers_23_input_layernorm_weight_to_fp16, y = out_93_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480117888))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480920768))))[name = string("lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [896]> lm_layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480922624)))];
            tensor<fp16, [4, 64, 896]> linear_161_cast_fp16 = linear(bias = lm_layers_23_self_attn_q_proj_bias_to_fp16, weight = lm_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_3631 = const()[name = string("op_3631"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 14, 64]> var_3632_cast_fp16 = reshape(shape = var_3631, x = linear_161_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [4]> q_perm_0 = const()[name = string("q_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480924480))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039232))))[name = string("lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_k_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039552)))];
            tensor<fp16, [4, 64, 128]> linear_162_cast_fp16 = linear(bias = lm_layers_23_self_attn_k_proj_bias_to_fp16, weight = lm_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3638_cast_fp16 = reshape(shape = var_3637, x = linear_162_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [128, 896]> lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [128, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481039872))), scale = tensor<fp16, [128, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154624))))[name = string("lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [128]> lm_layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = string("lm_layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481154944)))];
            tensor<fp16, [4, 64, 128]> linear_163_cast_fp16 = linear(bias = lm_layers_23_self_attn_v_proj_bias_to_fp16, weight = lm_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = hidden_states_323_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_3643 = const()[name = string("op_3643"), val = tensor<int32, [4]>([4, 64, -1, 64])];
            tensor<fp16, [4, 64, 2, 64]> var_3644_cast_fp16 = reshape(shape = var_3643, x = linear_163_cast_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> hidden_states_329_perm_0 = const()[name = string("hidden_states_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [4, 14, 64, 64]> q_cast_fp16 = transpose(perm = q_perm_0, x = var_3632_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [4, 14, 64, 64]> var_3648_cast_fp16 = mul(x = q_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([4, 14, 64, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 14, 64, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 14, 64, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_429_promoted_to_fp16 = const()[name = string("const_429_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 14, 64, 32]> var_3659_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_429_promoted_to_fp16)[name = string("op_3659_cast_fp16")];
            bool var_3661_interleave_0 = const()[name = string("op_3661_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> var_3661_cast_fp16 = concat(axis = var_20, interleave = var_3661_interleave_0, values = (var_3659_cast_fp16, x1_93_cast_fp16))[name = string("op_3661_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> var_3662_cast_fp16 = mul(x = var_3661_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> query_cast_fp16 = add(x = var_3648_cast_fp16, y = var_3662_cast_fp16)[name = string("query_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = var_3638_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [4, 2, 64, 64]> var_3664_cast_fp16 = mul(x = k_cast_fp16, y = cos_5_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([4, 2, 64, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 2, 64, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([4, 2, 64, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [4, 2, 64, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_432_promoted_to_fp16 = const()[name = string("const_432_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [4, 2, 64, 32]> var_3675_cast_fp16 = mul(x = x2_cast_fp16, y = const_432_promoted_to_fp16)[name = string("op_3675_cast_fp16")];
            bool var_3677_interleave_0 = const()[name = string("op_3677_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 2, 64, 64]> var_3677_cast_fp16 = concat(axis = var_20, interleave = var_3677_interleave_0, values = (var_3675_cast_fp16, x1_cast_fp16))[name = string("op_3677_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> var_3678_cast_fp16 = mul(x = var_3677_cast_fp16, y = sin_5_cast_fp16)[name = string("op_3678_cast_fp16")];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_325_cast_fp16 = add(x = var_3664_cast_fp16, y = var_3678_cast_fp16)[name = string("hidden_states_325_cast_fp16")];
            tensor<int32, [1]> var_3687_axes_0 = const()[name = string("op_3687_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3687_cast_fp16 = expand_dims(axes = var_3687_axes_0, x = hidden_states_325_cast_fp16)[name = string("op_3687_cast_fp16")];
            tensor<int32, [5]> hidden_states_327_reps_0 = const()[name = string("hidden_states_327_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_327_cast_fp16 = tile(reps = hidden_states_327_reps_0, x = var_3687_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> key_cast_fp16 = reshape(shape = var_3695, x = hidden_states_327_cast_fp16)[name = string("key_cast_fp16")];
            tensor<int32, [1]> var_3704_axes_0 = const()[name = string("op_3704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 2, 64, 64]> hidden_states_329_cast_fp16 = transpose(perm = hidden_states_329_perm_0, x = var_3644_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [4, 2, 1, 64, 64]> var_3704_cast_fp16 = expand_dims(axes = var_3704_axes_0, x = hidden_states_329_cast_fp16)[name = string("op_3704_cast_fp16")];
            tensor<int32, [5]> hidden_states_331_reps_0 = const()[name = string("hidden_states_331_reps_0"), val = tensor<int32, [5]>([1, 1, 7, 1, 1])];
            tensor<fp16, [4, 2, 7, 64, 64]> hidden_states_331_cast_fp16 = tile(reps = hidden_states_331_reps_0, x = var_3704_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<int32, [4]> var_3712 = const()[name = string("op_3712"), val = tensor<int32, [4]>([4, 14, 64, 64])];
            tensor<fp16, [4, 14, 64, 64]> value_cast_fp16 = reshape(shape = var_3712, x = hidden_states_331_cast_fp16)[name = string("value_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> mul_47_cast_fp16 = mul(x = query_cast_fp16, y = var_13_to_fp16)[name = string("mul_47_cast_fp16")];
            bool matmul_23_transpose_y_0 = const()[name = string("matmul_23_transpose_y_0"), val = bool(true)];
            bool matmul_23_transpose_x_0 = const()[name = string("matmul_23_transpose_x_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_0, transpose_y = matmul_23_transpose_y_0, x = mul_47_cast_fp16, y = key_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<fp16, [4, 14, 64, 64]> add_23_cast_fp16 = add(x = matmul_23_cast_fp16, y = mul_0_to_fp16_quantized)[name = string("add_23_cast_fp16")];
            int32 softmax_23_axis_0 = const()[name = string("softmax_23_axis_0"), val = int32(-1)];
            tensor<fp16, [4, 14, 64, 64]> softmax_23_cast_fp16 = softmax(axis = softmax_23_axis_0, x = add_23_cast_fp16)[name = string("softmax_23_cast_fp16")];
            bool attn_output_93_transpose_x_0 = const()[name = string("attn_output_93_transpose_x_0"), val = bool(false)];
            bool attn_output_93_transpose_y_0 = const()[name = string("attn_output_93_transpose_y_0"), val = bool(false)];
            tensor<fp16, [4, 14, 64, 64]> attn_output_93_cast_fp16 = matmul(transpose_x = attn_output_93_transpose_x_0, transpose_y = attn_output_93_transpose_y_0, x = softmax_23_cast_fp16, y = value_cast_fp16)[name = string("attn_output_93_cast_fp16")];
            tensor<int32, [4]> var_3715_perm_0 = const()[name = string("op_3715_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [3]>([4, 64, -1])];
            tensor<fp16, [4, 64, 14, 64]> var_3715_cast_fp16 = transpose(perm = var_3715_perm_0, x = attn_output_93_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [4, 64, 896]> var_3718_cast_fp16 = reshape(shape = var_3717, x = var_3715_cast_fp16)[name = string("op_3718_cast_fp16")];
            tensor<fp16, [896, 896]> lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481155264))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481958144))))[name = string("lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_164_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = var_3718_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_93_cast_fp16 = add(x = x_91_cast_fp16, y = linear_164_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3724_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3724_cast_fp16")];
            tensor<int32, [1]> var_3726_axes_0 = const()[name = string("op_3726_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3726_cast_fp16 = reduce_max(axes = var_3726_axes_0, keep_dims = var_12, x = var_3724_cast_fp16)[name = string("op_3726_cast_fp16")];
            fp16 const_441_to_fp16 = const()[name = string("const_441_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_47_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_441_to_fp16, x = var_3726_cast_fp16)[name = string("clip_47_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_95_cast_fp16 = real_div(x = x_93_cast_fp16, y = clip_47_cast_fp16)[name = string("xn_95_cast_fp16")];
            fp16 var_23_promoted_47_to_fp16 = const()[name = string("op_23_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3729_cast_fp16 = pow(x = xn_95_cast_fp16, y = var_23_promoted_47_to_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_3729_cast_fp16)[name = string("var_95_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3732_cast_fp16 = mul(x = clip_47_cast_fp16, y = clip_47_cast_fp16)[name = string("op_3732_cast_fp16")];
            fp32 var_3733_epsilon_0 = const()[name = string("op_3733_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3733_cast_fp16 = inverse(epsilon = var_3733_epsilon_0, x = var_3732_cast_fp16)[name = string("op_3733_cast_fp16")];
            fp16 var_3734_to_fp16 = const()[name = string("op_3734_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3735_cast_fp16 = mul(x = var_3733_cast_fp16, y = var_3734_to_fp16)[name = string("op_3735_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3736_cast_fp16 = add(x = var_95_cast_fp16, y = var_3735_cast_fp16)[name = string("op_3736_cast_fp16")];
            fp32 var_3737_epsilon_0 = const()[name = string("op_3737_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3737_cast_fp16 = rsqrt(epsilon = var_3737_epsilon_0, x = var_3736_cast_fp16)[name = string("op_3737_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_95_cast_fp16 = mul(x = xn_95_cast_fp16, y = var_3737_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [896]> lm_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("lm_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481960000)))];
            tensor<fp16, [4, 64, 896]> input_187_cast_fp16 = mul(x = lm_layers_23_post_attention_layernorm_weight_to_fp16, y = out_95_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481961856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486320064))))[name = string("lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> var_3745_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [4864, 896]> lm_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4864, 896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486329856))), scale = tensor<fp16, [4864, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490688064))))[name = string("lm_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 4864]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = lm_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_187_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [4, 64, 4864]> input_191_cast_fp16 = mul(x = var_3745_cast_fp16, y = linear_166_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [896, 4864]> lm_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [896, 4864]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490697856))), scale = tensor<fp16, [896, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495056064))))[name = string("lm_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [4, 64, 896]> linear_167_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = lm_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [4, 64, 896]> x_cast_fp16 = add(x = x_93_cast_fp16, y = linear_167_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3753_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_3753_cast_fp16")];
            tensor<int32, [1]> var_3755_axes_0 = const()[name = string("op_3755_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 64, 1]> var_3755_cast_fp16 = reduce_max(axes = var_3755_axes_0, keep_dims = var_12, x = var_3753_cast_fp16)[name = string("op_3755_cast_fp16")];
            fp16 const_442_to_fp16 = const()[name = string("const_442_to_fp16"), val = fp16(inf)];
            tensor<fp16, [4, 64, 1]> clip_48_cast_fp16 = clip(alpha = var_11_to_fp16, beta = const_442_to_fp16, x = var_3755_cast_fp16)[name = string("clip_48_cast_fp16")];
            tensor<fp16, [4, 64, 896]> xn_cast_fp16 = real_div(x = x_cast_fp16, y = clip_48_cast_fp16)[name = string("xn_cast_fp16")];
            fp16 var_23_promoted_48_to_fp16 = const()[name = string("op_23_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 64, 896]> var_3758_cast_fp16 = pow(x = xn_cast_fp16, y = var_23_promoted_48_to_fp16)[name = string("op_3758_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [4, 64, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_3758_cast_fp16)[name = string("var_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3761_cast_fp16 = mul(x = clip_48_cast_fp16, y = clip_48_cast_fp16)[name = string("op_3761_cast_fp16")];
            fp32 var_3762_epsilon_0 = const()[name = string("op_3762_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [4, 64, 1]> var_3762_cast_fp16 = inverse(epsilon = var_3762_epsilon_0, x = var_3761_cast_fp16)[name = string("op_3762_cast_fp16")];
            fp16 var_3763_to_fp16 = const()[name = string("op_3763_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [4, 64, 1]> var_3764_cast_fp16 = mul(x = var_3762_cast_fp16, y = var_3763_to_fp16)[name = string("op_3764_cast_fp16")];
            tensor<fp16, [4, 64, 1]> var_3765_cast_fp16 = add(x = var_cast_fp16, y = var_3764_cast_fp16)[name = string("op_3765_cast_fp16")];
            fp32 var_3766_epsilon_0 = const()[name = string("op_3766_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [4, 64, 1]> var_3766_cast_fp16 = rsqrt(epsilon = var_3766_epsilon_0, x = var_3765_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<fp16, [4, 64, 896]> out_cast_fp16 = mul(x = xn_cast_fp16, y = var_3766_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [896]> lm_norm_weight_to_fp16 = const()[name = string("lm_norm_weight_to_fp16"), val = tensor<fp16, [896]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495057920)))];
            tensor<fp16, [4, 64, 896]> hidden_cast_fp16 = mul(x = lm_norm_weight_to_fp16, y = out_cast_fp16)[name = string("hidden_cast_fp16")];
            tensor<int32, [1]> var_3770_axes_0 = const()[name = string("op_3770_axes_0"), val = tensor<int32, [1]>([-1])];
            string selector_to_fp16_dtype_0 = const()[name = string("selector_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 64]> selector_to_fp16 = cast(dtype = selector_to_fp16_dtype_0, x = selector)[name = string("cast_1")];
            tensor<fp16, [4, 64, 1]> var_3770_cast_fp16 = expand_dims(axes = var_3770_axes_0, x = selector_to_fp16)[name = string("op_3770_cast_fp16")];
            tensor<fp16, [4, 64, 896]> var_3771_cast_fp16 = mul(x = var_3770_cast_fp16, y = hidden_cast_fp16)[name = string("op_3771_cast_fp16")];
            tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([1])];
            bool input_keep_dims_0 = const()[name = string("input_keep_dims_0"), val = bool(false)];
            tensor<fp16, [4, 896]> input_cast_fp16 = reduce_sum(axes = input_axes_0, keep_dims = input_keep_dims_0, x = var_3771_cast_fp16)[name = string("input_cast_fp16")];
            tensor<int32, [1]> var_3779 = const()[name = string("op_3779"), val = tensor<int32, [1]>([-1])];
            bool var_3780 = const()[name = string("op_3780"), val = bool(true)];
            tensor<fp16, [4, 1]> var_3782_cast_fp16 = reduce_l2_norm(axes = var_3779, keep_dims = var_3780, x = input_cast_fp16)[name = string("op_3782_cast_fp16")];
            fp16 var_3783_to_fp16 = const()[name = string("op_3783_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [4, 1]> var_3784_cast_fp16 = maximum(x = var_3782_cast_fp16, y = var_3783_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<int32, [2]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [2]>([1, 896])];
            tensor<fp16, [4, 896]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_3784_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [4, 896]> var_3786_cast_fp16 = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = string("op_3786_cast_fp16")];
            string var_3786_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_3786_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp32, [4, 896]> embedding = cast(dtype = var_3786_cast_fp16_to_fp32_dtype_0, x = var_3786_cast_fp16)[name = string("cast_0")];
        } -> (embedding);
}