program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func main<ios18>(tensor<fp32, [1, 1, 128, 1024]> attention_mask, tensor<fp32, [1, 128, 1024]> input_embeds, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> k_cache_9, tensor<int32, [128]> positions, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_0, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_1, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_10, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_11, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_12, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_13, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_2, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_3, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_4, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_5, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_6, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_7, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_8, state<tensor<fp16, [1, 8, 1024, 128]>> v_cache_9) {
            int32 var_66_one_hot_vector_size_0 = const()[name = string("op_66_one_hot_vector_size_0"), val = int32(1024)];
            int32 var_66_axis_0 = const()[name = string("op_66_axis_0"), val = int32(-1)];
            int32 var_66_on_value_0 = const()[name = string("op_66_on_value_0"), val = int32(1)];
            int32 var_66_off_value_0 = const()[name = string("op_66_off_value_0"), val = int32(0)];
            tensor<int32, [128, 1024]> var_66 = one_hot(axis = var_66_axis_0, indices = positions, off_value = var_66_off_value_0, on_value = var_66_on_value_0, one_hot_vector_size = var_66_one_hot_vector_size_0)[name = string("op_66")];
            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([0])];
            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
            string cast_1_to_fp16_dtype_0 = const()[name = string("cast_1_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [128, 1024]> var_66_to_fp16 = cast(dtype = cast_1_to_fp16_dtype_0, x = var_66)[name = string("cast_3")];
            tensor<fp16, [1024]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_66_to_fp16)[name = string("op_76_cast_fp16")];
            tensor<int32, [4]> var_81 = const()[name = string("op_81"), val = tensor<int32, [4]>([1, 1, 1024, 1])];
            tensor<fp16, [1, 1, 1024, 1]> var_82_cast_fp16 = reshape(shape = var_81, x = var_76_cast_fp16)[name = string("op_82_cast_fp16")];
            int32 var_97 = const()[name = string("op_97"), val = int32(-1)];
            string input_embeds_to_fp16_dtype_0 = const()[name = string("input_embeds_to_fp16_dtype_0"), val = string("fp16")];
            fp16 var_96_promoted_to_fp16 = const()[name = string("op_96_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> input_embeds_to_fp16 = cast(dtype = input_embeds_to_fp16_dtype_0, x = input_embeds)[name = string("cast_2")];
            tensor<fp16, [1, 128, 1024]> var_106_cast_fp16 = pow(x = input_embeds_to_fp16, y = var_96_promoted_to_fp16)[name = string("op_106_cast_fp16")];
            tensor<int32, [1]> var_108_axes_0 = const()[name = string("op_108_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_108_keep_dims_0 = const()[name = string("op_108_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_108_cast_fp16 = reduce_mean(axes = var_108_axes_0, keep_dims = var_108_keep_dims_0, x = var_106_cast_fp16)[name = string("op_108_cast_fp16")];
            fp16 var_109_to_fp16 = const()[name = string("op_109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_110_cast_fp16 = add(x = var_108_cast_fp16, y = var_109_to_fp16)[name = string("op_110_cast_fp16")];
            fp32 norm_1_epsilon_0 = const()[name = string("norm_1_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_1_cast_fp16 = rsqrt(epsilon = norm_1_epsilon_0, x = var_110_cast_fp16)[name = string("norm_1_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_112_cast_fp16 = mul(x = input_embeds_to_fp16, y = norm_1_cast_fp16)[name = string("op_112_cast_fp16")];
            tensor<fp16, [1024]> layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [1, 128, 1024]> var_113_cast_fp16 = mul(x = var_112_cast_fp16, y = layers_0_input_layernorm_weight_to_fp16)[name = string("op_113_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2176))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099392))))[name = string("layers_0_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2099968)))];
            tensor<fp16, [1, 128, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = var_113_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_129 = const()[name = string("op_129"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_130_cast_fp16 = reshape(shape = var_129, x = linear_0_cast_fp16)[name = string("op_130_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2104128))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3152768))))[name = string("layers_0_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3153344)))];
            tensor<fp16, [1, 128, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = var_113_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_134 = const()[name = string("op_134"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_135_cast_fp16 = reshape(shape = var_134, x = linear_1_cast_fp16)[name = string("op_135_cast_fp16")];
            tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3155456))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204096))))[name = string("layers_0_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = var_113_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_139 = const()[name = string("op_139"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_140_cast_fp16 = reshape(shape = var_139, x = linear_2_cast_fp16)[name = string("op_140_cast_fp16")];
            tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_96_promoted_1_to_fp16 = const()[name = string("op_96_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_130_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 128, 128]> var_144_cast_fp16 = pow(x = x_5_cast_fp16, y = var_96_promoted_1_to_fp16)[name = string("op_144_cast_fp16")];
            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_146_keep_dims_0 = const()[name = string("op_146_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_146_cast_fp16 = reduce_mean(axes = var_146_axes_0, keep_dims = var_146_keep_dims_0, x = var_144_cast_fp16)[name = string("op_146_cast_fp16")];
            fp16 var_147_to_fp16 = const()[name = string("op_147_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_148_cast_fp16 = add(x = var_146_cast_fp16, y = var_147_to_fp16)[name = string("op_148_cast_fp16")];
            fp32 norm_3_epsilon_0 = const()[name = string("norm_3_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_3_cast_fp16 = rsqrt(epsilon = norm_3_epsilon_0, x = var_148_cast_fp16)[name = string("norm_3_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_150_cast_fp16 = mul(x = x_5_cast_fp16, y = norm_3_cast_fp16)[name = string("op_150_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204672)))];
            tensor<fp16, [1, 16, 128, 128]> var_151_cast_fp16 = mul(x = var_150_cast_fp16, y = layers_0_self_attn_q_norm_weight_to_fp16)[name = string("op_151_cast_fp16")];
            fp16 var_96_promoted_2_to_fp16 = const()[name = string("op_96_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_135_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 128, 128]> var_155_cast_fp16 = pow(x = x_9_cast_fp16, y = var_96_promoted_2_to_fp16)[name = string("op_155_cast_fp16")];
            tensor<int32, [1]> var_157_axes_0 = const()[name = string("op_157_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_157_keep_dims_0 = const()[name = string("op_157_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_157_cast_fp16 = reduce_mean(axes = var_157_axes_0, keep_dims = var_157_keep_dims_0, x = var_155_cast_fp16)[name = string("op_157_cast_fp16")];
            fp16 var_158_to_fp16 = const()[name = string("op_158_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_159_cast_fp16 = add(x = var_157_cast_fp16, y = var_158_to_fp16)[name = string("op_159_cast_fp16")];
            fp32 norm_5_epsilon_0 = const()[name = string("norm_5_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_5_cast_fp16 = rsqrt(epsilon = norm_5_epsilon_0, x = var_159_cast_fp16)[name = string("norm_5_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_161_cast_fp16 = mul(x = x_9_cast_fp16, y = norm_5_cast_fp16)[name = string("op_161_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4204992)))];
            tensor<fp16, [1, 8, 128, 128]> var_162_cast_fp16 = mul(x = var_161_cast_fp16, y = layers_0_self_attn_k_norm_weight_to_fp16)[name = string("op_162_cast_fp16")];
            tensor<int32, [1]> var_166_axes_0 = const()[name = string("op_166_axes_0"), val = tensor<int32, [1]>([-1])];
            string cast_12_to_fp16_dtype_0 = const()[name = string("cast_12_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [128]> positions_to_fp16 = cast(dtype = cast_12_to_fp16_dtype_0, x = positions)[name = string("cast_1")];
            tensor<fp16, [128, 1]> var_166_cast_fp16 = expand_dims(axes = var_166_axes_0, x = positions_to_fp16)[name = string("op_166_cast_fp16")];
            tensor<fp16, [64]> layers_0_self_attn_rope_inv_freq_to_fp16 = const()[name = string("layers_0_self_attn_rope_inv_freq_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4205312)))];
            tensor<fp16, [128, 64]> freqs_1_cast_fp16 = mul(x = var_166_cast_fp16, y = layers_0_self_attn_rope_inv_freq_to_fp16)[name = string("freqs_1_cast_fp16")];
            tensor<fp16, [128, 64]> var_168_cast_fp16 = cos(x = freqs_1_cast_fp16)[name = string("op_168_cast_fp16")];
            tensor<int32, [4]> var_170 = const()[name = string("op_170"), val = tensor<int32, [4]>([1, 1, -1, 64])];
            tensor<fp16, [1, 1, 128, 64]> cos_val_1_cast_fp16 = reshape(shape = var_170, x = var_168_cast_fp16)[name = string("cos_val_1_cast_fp16")];
            tensor<fp16, [128, 64]> var_172_cast_fp16 = sin(x = freqs_1_cast_fp16)[name = string("op_172_cast_fp16")];
            tensor<int32, [4]> var_174 = const()[name = string("op_174"), val = tensor<int32, [4]>([1, 1, -1, 64])];
            tensor<fp16, [1, 1, 128, 64]> sin_val_1_cast_fp16 = reshape(shape = var_174, x = var_172_cast_fp16)[name = string("sin_val_1_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_151_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, 64])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_151_cast_fp16)[name = string("x2_1_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_183_cast_fp16 = mul(x = x1_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_183_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_184_cast_fp16 = mul(x = x2_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_184_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_185_cast_fp16 = sub(x = var_183_cast_fp16, y = var_184_cast_fp16)[name = string("op_185_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_186_cast_fp16 = mul(x = x2_1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_186_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_187_cast_fp16 = mul(x = x1_1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_187_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_188_cast_fp16 = add(x = var_186_cast_fp16, y = var_187_cast_fp16)[name = string("op_188_cast_fp16")];
            bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_1_cast_fp16 = concat(axis = var_97, interleave = q_1_interleave_0, values = (var_185_cast_fp16, var_188_cast_fp16))[name = string("q_1_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_162_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, 64])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_162_cast_fp16)[name = string("x2_3_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_210_cast_fp16 = mul(x = x1_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_210_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_211_cast_fp16 = mul(x = x2_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_211_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_212_cast_fp16 = sub(x = var_210_cast_fp16, y = var_211_cast_fp16)[name = string("op_212_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_213_cast_fp16 = mul(x = x2_3_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_213_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_214_cast_fp16 = mul(x = x1_3_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_214_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_215_cast_fp16 = add(x = var_213_cast_fp16, y = var_214_cast_fp16)[name = string("op_215_cast_fp16")];
            bool var_217_interleave_0 = const()[name = string("op_217_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_217_cast_fp16 = concat(axis = var_97, interleave = var_217_interleave_0, values = (var_212_cast_fp16, var_215_cast_fp16))[name = string("op_217_cast_fp16")];
            tensor<int32, [4]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_4 = const()[name = string("concat_4"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = var_217_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [128, 1024]> reshape_1_cast_fp16 = reshape(shape = concat_4, x = transpose_1_cast_fp16)[name = string("reshape_1_cast_fp16")];
            bool matmul_0_transpose_x_1 = const()[name = string("matmul_0_transpose_x_1"), val = bool(true)];
            bool matmul_0_transpose_y_1 = const()[name = string("matmul_0_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_1, transpose_y = matmul_0_transpose_y_1, x = var_66_to_fp16, y = reshape_1_cast_fp16)[name = string("matmul_0_cast_fp16")];
            tensor<int32, [4]> concat_7 = const()[name = string("concat_7"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_2_cast_fp16 = reshape(shape = concat_7, x = matmul_0_cast_fp16)[name = string("reshape_2_cast_fp16")];
            tensor<int32, [4]> scattered_k_1_perm_0 = const()[name = string("scattered_k_1_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_12 = const()[name = string("concat_12"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_56_cast_fp16 = transpose(perm = transpose_56_perm_0, x = var_140_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [128, 1024]> reshape_4_cast_fp16 = reshape(shape = concat_12, x = transpose_56_cast_fp16)[name = string("reshape_4_cast_fp16")];
            bool matmul_1_transpose_x_1 = const()[name = string("matmul_1_transpose_x_1"), val = bool(true)];
            bool matmul_1_transpose_y_1 = const()[name = string("matmul_1_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_1, transpose_y = matmul_1_transpose_y_1, x = var_66_to_fp16, y = reshape_4_cast_fp16)[name = string("matmul_1_cast_fp16")];
            tensor<int32, [4]> concat_15 = const()[name = string("concat_15"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_5_cast_fp16 = reshape(shape = concat_15, x = matmul_1_cast_fp16)[name = string("reshape_5_cast_fp16")];
            tensor<int32, [4]> scattered_v_1_perm_0 = const()[name = string("scattered_v_1_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            fp16 var_99_promoted_to_fp16 = const()[name = string("op_99_promoted_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1, 1024, 1]> var_222_cast_fp16 = sub(x = var_99_promoted_to_fp16, y = var_82_cast_fp16)[name = string("op_222_cast_fp16")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_0 = read_state(input = k_cache_0)[name = string("read_state_0")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_3_cast_fp16 = mul(x = read_state_0, y = var_222_cast_fp16)[name = string("k_cache_3_cast_fp16")];
            write_state(data = k_cache_3_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_56_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_56 = read_state(input = k_cache_0)[name = string("coreml_update_state_56")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_1_cast_fp16 = transpose(perm = scattered_k_1_perm_0, x = reshape_2_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_5_cast_fp16 = add(x = coreml_update_state_56, y = scattered_k_1_cast_fp16)[name = string("k_cache_5_cast_fp16")];
            write_state(data = k_cache_5_cast_fp16, input = k_cache_0)[name = string("coreml_update_state_57_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_57 = read_state(input = k_cache_0)[name = string("coreml_update_state_57")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_1 = read_state(input = v_cache_0)[name = string("read_state_1")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_3_cast_fp16 = mul(x = read_state_1, y = var_222_cast_fp16)[name = string("v_cache_3_cast_fp16")];
            write_state(data = v_cache_3_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_58_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_58 = read_state(input = v_cache_0)[name = string("coreml_update_state_58")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_1_cast_fp16 = transpose(perm = scattered_v_1_perm_0, x = reshape_5_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_5_cast_fp16 = add(x = coreml_update_state_58, y = scattered_v_1_cast_fp16)[name = string("v_cache_5_cast_fp16")];
            write_state(data = v_cache_5_cast_fp16, input = v_cache_0)[name = string("coreml_update_state_59_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_59 = read_state(input = v_cache_0)[name = string("coreml_update_state_59")];
            tensor<int32, [1]> var_228_axes_0 = const()[name = string("op_228_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_228_cast_fp16 = expand_dims(axes = var_228_axes_0, x = coreml_update_state_57)[name = string("op_228_cast_fp16")];
            tensor<int32, [5]> k_exp_1_reps_0 = const()[name = string("k_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_1_cast_fp16 = tile(reps = k_exp_1_reps_0, x = var_228_cast_fp16)[name = string("k_exp_1_cast_fp16")];
            tensor<int32, [4]> var_231 = const()[name = string("op_231"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_3_cast_fp16 = reshape(shape = var_231, x = k_exp_1_cast_fp16)[name = string("k_exp_3_cast_fp16")];
            tensor<int32, [1]> var_233_axes_0 = const()[name = string("op_233_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_233_cast_fp16 = expand_dims(axes = var_233_axes_0, x = coreml_update_state_59)[name = string("op_233_cast_fp16")];
            tensor<int32, [5]> v_exp_1_reps_0 = const()[name = string("v_exp_1_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_1_cast_fp16 = tile(reps = v_exp_1_reps_0, x = var_233_cast_fp16)[name = string("v_exp_1_cast_fp16")];
            tensor<int32, [4]> var_236 = const()[name = string("op_236"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_3_cast_fp16 = reshape(shape = var_236, x = v_exp_1_cast_fp16)[name = string("v_exp_3_cast_fp16")];
            bool var_239_transpose_x_1 = const()[name = string("op_239_transpose_x_1"), val = bool(false)];
            bool var_239_transpose_y_1 = const()[name = string("op_239_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_239_cast_fp16 = matmul(transpose_x = var_239_transpose_x_1, transpose_y = var_239_transpose_y_1, x = q_1_cast_fp16, y = k_exp_3_cast_fp16)[name = string("op_239_cast_fp16")];
            fp16 var_240_to_fp16 = const()[name = string("op_240_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_1_cast_fp16 = mul(x = var_239_cast_fp16, y = var_240_to_fp16)[name = string("attn_1_cast_fp16")];
            string attention_mask_to_fp16_dtype_0 = const()[name = string("attention_mask_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 128, 1024]> attention_mask_to_fp16 = cast(dtype = attention_mask_to_fp16_dtype_0, x = attention_mask)[name = string("cast_0")];
            tensor<fp16, [1, 16, 128, 1024]> input_1_cast_fp16 = add(x = attn_1_cast_fp16, y = attention_mask_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_3_cast_fp16 = softmax(axis = var_97, x = input_1_cast_fp16)[name = string("attn_3_cast_fp16")];
            bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)];
            bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3_cast_fp16, y = v_exp_3_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<int32, [4]> var_245_perm_0 = const()[name = string("op_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_246 = const()[name = string("op_246"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_245_cast_fp16 = transpose(perm = var_245_perm_0, x = out_1_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 128, 2048]> input_3_cast_fp16 = reshape(shape = var_246, x = var_245_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_0_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4205504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6302720))))[name = string("layers_0_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_19_cast_fp16 = add(x = input_embeds_to_fp16, y = linear_3_cast_fp16)[name = string("x_19_cast_fp16")];
            fp16 var_96_promoted_3_to_fp16 = const()[name = string("op_96_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_253_cast_fp16 = pow(x = x_19_cast_fp16, y = var_96_promoted_3_to_fp16)[name = string("op_253_cast_fp16")];
            tensor<int32, [1]> var_255_axes_0 = const()[name = string("op_255_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_255_keep_dims_0 = const()[name = string("op_255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_255_cast_fp16 = reduce_mean(axes = var_255_axes_0, keep_dims = var_255_keep_dims_0, x = var_253_cast_fp16)[name = string("op_255_cast_fp16")];
            fp16 var_256_to_fp16 = const()[name = string("op_256_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_257_cast_fp16 = add(x = var_255_cast_fp16, y = var_256_to_fp16)[name = string("op_257_cast_fp16")];
            fp32 norm_7_epsilon_0 = const()[name = string("norm_7_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_7_cast_fp16 = rsqrt(epsilon = norm_7_epsilon_0, x = var_257_cast_fp16)[name = string("norm_7_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_259_cast_fp16 = mul(x = x_19_cast_fp16, y = norm_7_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<fp16, [1024]> layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6303296)))];
            tensor<fp16, [1, 128, 1024]> var_260_cast_fp16 = mul(x = var_259_cast_fp16, y = layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_260_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_0_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6305408))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9451200))))[name = string("layers_0_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9451776)))];
            tensor<fp16, [1, 128, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16_palettized, x = var_260_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_270_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_270_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_0_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9457984))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12603776))))[name = string("layers_0_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16_palettized, x = var_260_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_9_cast_fp16 = mul(x = var_270_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_0_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12604352))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15750144))))[name = string("layers_0_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_25_cast_fp16 = add(x = x_19_cast_fp16, y = linear_6_cast_fp16)[name = string("x_25_cast_fp16")];
            int32 var_291 = const()[name = string("op_291"), val = int32(-1)];
            fp16 var_290_promoted_to_fp16 = const()[name = string("op_290_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_300_cast_fp16 = pow(x = x_25_cast_fp16, y = var_290_promoted_to_fp16)[name = string("op_300_cast_fp16")];
            tensor<int32, [1]> var_302_axes_0 = const()[name = string("op_302_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_302_keep_dims_0 = const()[name = string("op_302_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_302_cast_fp16 = reduce_mean(axes = var_302_axes_0, keep_dims = var_302_keep_dims_0, x = var_300_cast_fp16)[name = string("op_302_cast_fp16")];
            fp16 var_303_to_fp16 = const()[name = string("op_303_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_304_cast_fp16 = add(x = var_302_cast_fp16, y = var_303_to_fp16)[name = string("op_304_cast_fp16")];
            fp32 norm_9_epsilon_0 = const()[name = string("norm_9_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_9_cast_fp16 = rsqrt(epsilon = norm_9_epsilon_0, x = var_304_cast_fp16)[name = string("norm_9_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_306_cast_fp16 = mul(x = x_25_cast_fp16, y = norm_9_cast_fp16)[name = string("op_306_cast_fp16")];
            tensor<fp16, [1024]> layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15750720)))];
            tensor<fp16, [1, 128, 1024]> var_307_cast_fp16 = mul(x = var_306_cast_fp16, y = layers_1_input_layernorm_weight_to_fp16)[name = string("op_307_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15752832))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17850048))))[name = string("layers_1_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = var_307_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_323 = const()[name = string("op_323"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_324_cast_fp16 = reshape(shape = var_323, x = linear_7_cast_fp16)[name = string("op_324_cast_fp16")];
            tensor<int32, [4]> x_31_perm_0 = const()[name = string("x_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17850624))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18899264))))[name = string("layers_1_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = var_307_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_328 = const()[name = string("op_328"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_329_cast_fp16 = reshape(shape = var_328, x = linear_8_cast_fp16)[name = string("op_329_cast_fp16")];
            tensor<int32, [4]> x_35_perm_0 = const()[name = string("x_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18899840))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19948480))))[name = string("layers_1_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = var_307_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_333 = const()[name = string("op_333"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_334_cast_fp16 = reshape(shape = var_333, x = linear_9_cast_fp16)[name = string("op_334_cast_fp16")];
            tensor<int32, [4]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_290_promoted_1_to_fp16 = const()[name = string("op_290_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_31_cast_fp16 = transpose(perm = x_31_perm_0, x = var_324_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 128, 128]> var_338_cast_fp16 = pow(x = x_31_cast_fp16, y = var_290_promoted_1_to_fp16)[name = string("op_338_cast_fp16")];
            tensor<int32, [1]> var_340_axes_0 = const()[name = string("op_340_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_340_keep_dims_0 = const()[name = string("op_340_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_340_cast_fp16 = reduce_mean(axes = var_340_axes_0, keep_dims = var_340_keep_dims_0, x = var_338_cast_fp16)[name = string("op_340_cast_fp16")];
            fp16 var_341_to_fp16 = const()[name = string("op_341_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_342_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_to_fp16)[name = string("op_342_cast_fp16")];
            fp32 norm_11_epsilon_0 = const()[name = string("norm_11_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_11_cast_fp16 = rsqrt(epsilon = norm_11_epsilon_0, x = var_342_cast_fp16)[name = string("norm_11_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_344_cast_fp16 = mul(x = x_31_cast_fp16, y = norm_11_cast_fp16)[name = string("op_344_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949056)))];
            tensor<fp16, [1, 16, 128, 128]> var_345_cast_fp16 = mul(x = var_344_cast_fp16, y = layers_1_self_attn_q_norm_weight_to_fp16)[name = string("op_345_cast_fp16")];
            fp16 var_290_promoted_2_to_fp16 = const()[name = string("op_290_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_35_cast_fp16 = transpose(perm = x_35_perm_0, x = var_329_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 8, 128, 128]> var_349_cast_fp16 = pow(x = x_35_cast_fp16, y = var_290_promoted_2_to_fp16)[name = string("op_349_cast_fp16")];
            tensor<int32, [1]> var_351_axes_0 = const()[name = string("op_351_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_351_keep_dims_0 = const()[name = string("op_351_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_351_cast_fp16 = reduce_mean(axes = var_351_axes_0, keep_dims = var_351_keep_dims_0, x = var_349_cast_fp16)[name = string("op_351_cast_fp16")];
            fp16 var_352_to_fp16 = const()[name = string("op_352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_353_cast_fp16 = add(x = var_351_cast_fp16, y = var_352_to_fp16)[name = string("op_353_cast_fp16")];
            fp32 norm_13_epsilon_0 = const()[name = string("norm_13_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_13_cast_fp16 = rsqrt(epsilon = norm_13_epsilon_0, x = var_353_cast_fp16)[name = string("norm_13_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_355_cast_fp16 = mul(x = x_35_cast_fp16, y = norm_13_cast_fp16)[name = string("op_355_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949376)))];
            tensor<fp16, [1, 8, 128, 128]> var_356_cast_fp16 = mul(x = var_355_cast_fp16, y = layers_1_self_attn_k_norm_weight_to_fp16)[name = string("op_356_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_345_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, 64])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_345_cast_fp16)[name = string("x2_5_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_377_cast_fp16 = mul(x = x1_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_377_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_378_cast_fp16 = mul(x = x2_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_378_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_379_cast_fp16 = sub(x = var_377_cast_fp16, y = var_378_cast_fp16)[name = string("op_379_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_380_cast_fp16 = mul(x = x2_5_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_380_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_381_cast_fp16 = mul(x = x1_5_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_381_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_382_cast_fp16 = add(x = var_380_cast_fp16, y = var_381_cast_fp16)[name = string("op_382_cast_fp16")];
            bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_3_cast_fp16 = concat(axis = var_291, interleave = q_3_interleave_0, values = (var_379_cast_fp16, var_382_cast_fp16))[name = string("q_3_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_356_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, 64])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_356_cast_fp16)[name = string("x2_7_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_404_cast_fp16 = mul(x = x1_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_404_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_405_cast_fp16 = mul(x = x2_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_405_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_406_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = string("op_406_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_407_cast_fp16 = mul(x = x2_7_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_407_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_408_cast_fp16 = mul(x = x1_7_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_408_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_409_cast_fp16 = add(x = var_407_cast_fp16, y = var_408_cast_fp16)[name = string("op_409_cast_fp16")];
            bool var_411_interleave_0 = const()[name = string("op_411_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_411_cast_fp16 = concat(axis = var_291, interleave = var_411_interleave_0, values = (var_406_cast_fp16, var_409_cast_fp16))[name = string("op_411_cast_fp16")];
            tensor<int32, [4]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_22 = const()[name = string("concat_22"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = var_411_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [128, 1024]> reshape_7_cast_fp16 = reshape(shape = concat_22, x = transpose_5_cast_fp16)[name = string("reshape_7_cast_fp16")];
            bool matmul_2_transpose_x_1 = const()[name = string("matmul_2_transpose_x_1"), val = bool(true)];
            bool matmul_2_transpose_y_1 = const()[name = string("matmul_2_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_1, transpose_y = matmul_2_transpose_y_1, x = var_66_to_fp16, y = reshape_7_cast_fp16)[name = string("matmul_2_cast_fp16")];
            tensor<int32, [4]> concat_25 = const()[name = string("concat_25"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_8_cast_fp16 = reshape(shape = concat_25, x = matmul_2_cast_fp16)[name = string("reshape_8_cast_fp16")];
            tensor<int32, [4]> scattered_k_3_perm_0 = const()[name = string("scattered_k_3_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_30 = const()[name = string("concat_30"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_57_cast_fp16 = transpose(perm = transpose_57_perm_0, x = var_334_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [128, 1024]> reshape_10_cast_fp16 = reshape(shape = concat_30, x = transpose_57_cast_fp16)[name = string("reshape_10_cast_fp16")];
            bool matmul_3_transpose_x_1 = const()[name = string("matmul_3_transpose_x_1"), val = bool(true)];
            bool matmul_3_transpose_y_1 = const()[name = string("matmul_3_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_1, transpose_y = matmul_3_transpose_y_1, x = var_66_to_fp16, y = reshape_10_cast_fp16)[name = string("matmul_3_cast_fp16")];
            tensor<int32, [4]> concat_33 = const()[name = string("concat_33"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_11_cast_fp16 = reshape(shape = concat_33, x = matmul_3_cast_fp16)[name = string("reshape_11_cast_fp16")];
            tensor<int32, [4]> scattered_v_3_perm_0 = const()[name = string("scattered_v_3_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_2 = read_state(input = k_cache_1)[name = string("read_state_2")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_9_cast_fp16 = mul(x = read_state_2, y = var_222_cast_fp16)[name = string("k_cache_9_cast_fp16")];
            write_state(data = k_cache_9_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_60_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_60 = read_state(input = k_cache_1)[name = string("coreml_update_state_60")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_3_cast_fp16 = transpose(perm = scattered_k_3_perm_0, x = reshape_8_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_11_cast_fp16 = add(x = coreml_update_state_60, y = scattered_k_3_cast_fp16)[name = string("k_cache_11_cast_fp16")];
            write_state(data = k_cache_11_cast_fp16, input = k_cache_1)[name = string("coreml_update_state_61_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_61 = read_state(input = k_cache_1)[name = string("coreml_update_state_61")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_3 = read_state(input = v_cache_1)[name = string("read_state_3")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_9_cast_fp16 = mul(x = read_state_3, y = var_222_cast_fp16)[name = string("v_cache_9_cast_fp16")];
            write_state(data = v_cache_9_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_62_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_62 = read_state(input = v_cache_1)[name = string("coreml_update_state_62")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_3_cast_fp16 = transpose(perm = scattered_v_3_perm_0, x = reshape_11_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_11_cast_fp16 = add(x = coreml_update_state_62, y = scattered_v_3_cast_fp16)[name = string("v_cache_11_cast_fp16")];
            write_state(data = v_cache_11_cast_fp16, input = v_cache_1)[name = string("coreml_update_state_63_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_63 = read_state(input = v_cache_1)[name = string("coreml_update_state_63")];
            tensor<int32, [1]> var_422_axes_0 = const()[name = string("op_422_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_422_cast_fp16 = expand_dims(axes = var_422_axes_0, x = coreml_update_state_61)[name = string("op_422_cast_fp16")];
            tensor<int32, [5]> k_exp_5_reps_0 = const()[name = string("k_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_5_cast_fp16 = tile(reps = k_exp_5_reps_0, x = var_422_cast_fp16)[name = string("k_exp_5_cast_fp16")];
            tensor<int32, [4]> var_425 = const()[name = string("op_425"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_7_cast_fp16 = reshape(shape = var_425, x = k_exp_5_cast_fp16)[name = string("k_exp_7_cast_fp16")];
            tensor<int32, [1]> var_427_axes_0 = const()[name = string("op_427_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_427_cast_fp16 = expand_dims(axes = var_427_axes_0, x = coreml_update_state_63)[name = string("op_427_cast_fp16")];
            tensor<int32, [5]> v_exp_5_reps_0 = const()[name = string("v_exp_5_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_5_cast_fp16 = tile(reps = v_exp_5_reps_0, x = var_427_cast_fp16)[name = string("v_exp_5_cast_fp16")];
            tensor<int32, [4]> var_430 = const()[name = string("op_430"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_7_cast_fp16 = reshape(shape = var_430, x = v_exp_5_cast_fp16)[name = string("v_exp_7_cast_fp16")];
            bool var_433_transpose_x_1 = const()[name = string("op_433_transpose_x_1"), val = bool(false)];
            bool var_433_transpose_y_1 = const()[name = string("op_433_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_433_cast_fp16 = matmul(transpose_x = var_433_transpose_x_1, transpose_y = var_433_transpose_y_1, x = q_3_cast_fp16, y = k_exp_7_cast_fp16)[name = string("op_433_cast_fp16")];
            fp16 var_434_to_fp16 = const()[name = string("op_434_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_5_cast_fp16 = mul(x = var_433_cast_fp16, y = var_434_to_fp16)[name = string("attn_5_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_11_cast_fp16 = add(x = attn_5_cast_fp16, y = attention_mask_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_7_cast_fp16 = softmax(axis = var_291, x = input_11_cast_fp16)[name = string("attn_7_cast_fp16")];
            bool out_3_transpose_x_0 = const()[name = string("out_3_transpose_x_0"), val = bool(false)];
            bool out_3_transpose_y_0 = const()[name = string("out_3_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_7_cast_fp16, y = v_exp_7_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<int32, [4]> var_439_perm_0 = const()[name = string("op_439_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_440 = const()[name = string("op_440"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_439_cast_fp16 = transpose(perm = var_439_perm_0, x = out_3_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 128, 2048]> input_13_cast_fp16 = reshape(shape = var_440, x = var_439_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_1_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19949696))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22046912))))[name = string("layers_1_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_45_cast_fp16 = add(x = x_25_cast_fp16, y = linear_10_cast_fp16)[name = string("x_45_cast_fp16")];
            fp16 var_290_promoted_3_to_fp16 = const()[name = string("op_290_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_447_cast_fp16 = pow(x = x_45_cast_fp16, y = var_290_promoted_3_to_fp16)[name = string("op_447_cast_fp16")];
            tensor<int32, [1]> var_449_axes_0 = const()[name = string("op_449_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_449_keep_dims_0 = const()[name = string("op_449_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_449_cast_fp16 = reduce_mean(axes = var_449_axes_0, keep_dims = var_449_keep_dims_0, x = var_447_cast_fp16)[name = string("op_449_cast_fp16")];
            fp16 var_450_to_fp16 = const()[name = string("op_450_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_451_cast_fp16 = add(x = var_449_cast_fp16, y = var_450_to_fp16)[name = string("op_451_cast_fp16")];
            fp32 norm_15_epsilon_0 = const()[name = string("norm_15_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_15_cast_fp16 = rsqrt(epsilon = norm_15_epsilon_0, x = var_451_cast_fp16)[name = string("norm_15_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_453_cast_fp16 = mul(x = x_45_cast_fp16, y = norm_15_cast_fp16)[name = string("op_453_cast_fp16")];
            tensor<fp16, [1024]> layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22047488)))];
            tensor<fp16, [1, 128, 1024]> var_454_cast_fp16 = mul(x = var_453_cast_fp16, y = layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_454_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_1_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22049600))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25195392))))[name = string("layers_1_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16_palettized, x = var_454_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_464_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_1_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25195968))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28341760))))[name = string("layers_1_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16_palettized, x = var_454_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_19_cast_fp16 = mul(x = var_464_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_1_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28342336))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31488128))))[name = string("layers_1_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_51_cast_fp16 = add(x = x_45_cast_fp16, y = linear_13_cast_fp16)[name = string("x_51_cast_fp16")];
            int32 var_485 = const()[name = string("op_485"), val = int32(-1)];
            fp16 var_484_promoted_to_fp16 = const()[name = string("op_484_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_494_cast_fp16 = pow(x = x_51_cast_fp16, y = var_484_promoted_to_fp16)[name = string("op_494_cast_fp16")];
            tensor<int32, [1]> var_496_axes_0 = const()[name = string("op_496_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_496_keep_dims_0 = const()[name = string("op_496_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_496_cast_fp16 = reduce_mean(axes = var_496_axes_0, keep_dims = var_496_keep_dims_0, x = var_494_cast_fp16)[name = string("op_496_cast_fp16")];
            fp16 var_497_to_fp16 = const()[name = string("op_497_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_498_cast_fp16 = add(x = var_496_cast_fp16, y = var_497_to_fp16)[name = string("op_498_cast_fp16")];
            fp32 norm_17_epsilon_0 = const()[name = string("norm_17_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_17_cast_fp16 = rsqrt(epsilon = norm_17_epsilon_0, x = var_498_cast_fp16)[name = string("norm_17_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_500_cast_fp16 = mul(x = x_51_cast_fp16, y = norm_17_cast_fp16)[name = string("op_500_cast_fp16")];
            tensor<fp16, [1024]> layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31488704)))];
            tensor<fp16, [1, 128, 1024]> var_501_cast_fp16 = mul(x = var_500_cast_fp16, y = layers_2_input_layernorm_weight_to_fp16)[name = string("op_501_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31490816))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33588032))))[name = string("layers_2_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = var_501_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_517 = const()[name = string("op_517"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_518_cast_fp16 = reshape(shape = var_517, x = linear_14_cast_fp16)[name = string("op_518_cast_fp16")];
            tensor<int32, [4]> x_57_perm_0 = const()[name = string("x_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33588608))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34637248))))[name = string("layers_2_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = var_501_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_522 = const()[name = string("op_522"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_523_cast_fp16 = reshape(shape = var_522, x = linear_15_cast_fp16)[name = string("op_523_cast_fp16")];
            tensor<int32, [4]> x_61_perm_0 = const()[name = string("x_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34637824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35686464))))[name = string("layers_2_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = var_501_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_527 = const()[name = string("op_527"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_528_cast_fp16 = reshape(shape = var_527, x = linear_16_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_484_promoted_1_to_fp16 = const()[name = string("op_484_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_57_cast_fp16 = transpose(perm = x_57_perm_0, x = var_518_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, 128, 128]> var_532_cast_fp16 = pow(x = x_57_cast_fp16, y = var_484_promoted_1_to_fp16)[name = string("op_532_cast_fp16")];
            tensor<int32, [1]> var_534_axes_0 = const()[name = string("op_534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_534_keep_dims_0 = const()[name = string("op_534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_534_cast_fp16 = reduce_mean(axes = var_534_axes_0, keep_dims = var_534_keep_dims_0, x = var_532_cast_fp16)[name = string("op_534_cast_fp16")];
            fp16 var_535_to_fp16 = const()[name = string("op_535_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_536_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_to_fp16)[name = string("op_536_cast_fp16")];
            fp32 norm_19_epsilon_0 = const()[name = string("norm_19_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_19_cast_fp16 = rsqrt(epsilon = norm_19_epsilon_0, x = var_536_cast_fp16)[name = string("norm_19_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_538_cast_fp16 = mul(x = x_57_cast_fp16, y = norm_19_cast_fp16)[name = string("op_538_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687040)))];
            tensor<fp16, [1, 16, 128, 128]> var_539_cast_fp16 = mul(x = var_538_cast_fp16, y = layers_2_self_attn_q_norm_weight_to_fp16)[name = string("op_539_cast_fp16")];
            fp16 var_484_promoted_2_to_fp16 = const()[name = string("op_484_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_61_cast_fp16 = transpose(perm = x_61_perm_0, x = var_523_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 128, 128]> var_543_cast_fp16 = pow(x = x_61_cast_fp16, y = var_484_promoted_2_to_fp16)[name = string("op_543_cast_fp16")];
            tensor<int32, [1]> var_545_axes_0 = const()[name = string("op_545_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_545_keep_dims_0 = const()[name = string("op_545_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_545_cast_fp16 = reduce_mean(axes = var_545_axes_0, keep_dims = var_545_keep_dims_0, x = var_543_cast_fp16)[name = string("op_545_cast_fp16")];
            fp16 var_546_to_fp16 = const()[name = string("op_546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_547_cast_fp16 = add(x = var_545_cast_fp16, y = var_546_to_fp16)[name = string("op_547_cast_fp16")];
            fp32 norm_21_epsilon_0 = const()[name = string("norm_21_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_21_cast_fp16 = rsqrt(epsilon = norm_21_epsilon_0, x = var_547_cast_fp16)[name = string("norm_21_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_549_cast_fp16 = mul(x = x_61_cast_fp16, y = norm_21_cast_fp16)[name = string("op_549_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687360)))];
            tensor<fp16, [1, 8, 128, 128]> var_550_cast_fp16 = mul(x = var_549_cast_fp16, y = layers_2_self_attn_k_norm_weight_to_fp16)[name = string("op_550_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_539_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, 64])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_539_cast_fp16)[name = string("x2_9_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_571_cast_fp16 = mul(x = x1_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_571_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_572_cast_fp16 = mul(x = x2_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_572_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_573_cast_fp16 = sub(x = var_571_cast_fp16, y = var_572_cast_fp16)[name = string("op_573_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_574_cast_fp16 = mul(x = x2_9_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_574_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_575_cast_fp16 = mul(x = x1_9_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_575_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_576_cast_fp16 = add(x = var_574_cast_fp16, y = var_575_cast_fp16)[name = string("op_576_cast_fp16")];
            bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_5_cast_fp16 = concat(axis = var_485, interleave = q_5_interleave_0, values = (var_573_cast_fp16, var_576_cast_fp16))[name = string("q_5_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_550_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, 64])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_550_cast_fp16)[name = string("x2_11_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_598_cast_fp16 = mul(x = x1_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_598_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_599_cast_fp16 = mul(x = x2_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_599_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_600_cast_fp16 = sub(x = var_598_cast_fp16, y = var_599_cast_fp16)[name = string("op_600_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_601_cast_fp16 = mul(x = x2_11_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_601_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_602_cast_fp16 = mul(x = x1_11_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_602_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_603_cast_fp16 = add(x = var_601_cast_fp16, y = var_602_cast_fp16)[name = string("op_603_cast_fp16")];
            bool var_605_interleave_0 = const()[name = string("op_605_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_605_cast_fp16 = concat(axis = var_485, interleave = var_605_interleave_0, values = (var_600_cast_fp16, var_603_cast_fp16))[name = string("op_605_cast_fp16")];
            tensor<int32, [4]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_40 = const()[name = string("concat_40"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = var_605_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [128, 1024]> reshape_13_cast_fp16 = reshape(shape = concat_40, x = transpose_9_cast_fp16)[name = string("reshape_13_cast_fp16")];
            bool matmul_4_transpose_x_1 = const()[name = string("matmul_4_transpose_x_1"), val = bool(true)];
            bool matmul_4_transpose_y_1 = const()[name = string("matmul_4_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_1, transpose_y = matmul_4_transpose_y_1, x = var_66_to_fp16, y = reshape_13_cast_fp16)[name = string("matmul_4_cast_fp16")];
            tensor<int32, [4]> concat_43 = const()[name = string("concat_43"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_14_cast_fp16 = reshape(shape = concat_43, x = matmul_4_cast_fp16)[name = string("reshape_14_cast_fp16")];
            tensor<int32, [4]> scattered_k_5_perm_0 = const()[name = string("scattered_k_5_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_48 = const()[name = string("concat_48"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_58_cast_fp16 = transpose(perm = transpose_58_perm_0, x = var_528_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [128, 1024]> reshape_16_cast_fp16 = reshape(shape = concat_48, x = transpose_58_cast_fp16)[name = string("reshape_16_cast_fp16")];
            bool matmul_5_transpose_x_1 = const()[name = string("matmul_5_transpose_x_1"), val = bool(true)];
            bool matmul_5_transpose_y_1 = const()[name = string("matmul_5_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_1, transpose_y = matmul_5_transpose_y_1, x = var_66_to_fp16, y = reshape_16_cast_fp16)[name = string("matmul_5_cast_fp16")];
            tensor<int32, [4]> concat_51 = const()[name = string("concat_51"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_17_cast_fp16 = reshape(shape = concat_51, x = matmul_5_cast_fp16)[name = string("reshape_17_cast_fp16")];
            tensor<int32, [4]> scattered_v_5_perm_0 = const()[name = string("scattered_v_5_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_4 = read_state(input = k_cache_2)[name = string("read_state_4")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_15_cast_fp16 = mul(x = read_state_4, y = var_222_cast_fp16)[name = string("k_cache_15_cast_fp16")];
            write_state(data = k_cache_15_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_64_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_64 = read_state(input = k_cache_2)[name = string("coreml_update_state_64")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_5_cast_fp16 = transpose(perm = scattered_k_5_perm_0, x = reshape_14_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_17_cast_fp16 = add(x = coreml_update_state_64, y = scattered_k_5_cast_fp16)[name = string("k_cache_17_cast_fp16")];
            write_state(data = k_cache_17_cast_fp16, input = k_cache_2)[name = string("coreml_update_state_65_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_65 = read_state(input = k_cache_2)[name = string("coreml_update_state_65")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_5 = read_state(input = v_cache_2)[name = string("read_state_5")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_15_cast_fp16 = mul(x = read_state_5, y = var_222_cast_fp16)[name = string("v_cache_15_cast_fp16")];
            write_state(data = v_cache_15_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_66_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_66 = read_state(input = v_cache_2)[name = string("coreml_update_state_66")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_5_cast_fp16 = transpose(perm = scattered_v_5_perm_0, x = reshape_17_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_17_cast_fp16 = add(x = coreml_update_state_66, y = scattered_v_5_cast_fp16)[name = string("v_cache_17_cast_fp16")];
            write_state(data = v_cache_17_cast_fp16, input = v_cache_2)[name = string("coreml_update_state_67_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_67 = read_state(input = v_cache_2)[name = string("coreml_update_state_67")];
            tensor<int32, [1]> var_616_axes_0 = const()[name = string("op_616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_616_cast_fp16 = expand_dims(axes = var_616_axes_0, x = coreml_update_state_65)[name = string("op_616_cast_fp16")];
            tensor<int32, [5]> k_exp_9_reps_0 = const()[name = string("k_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_9_cast_fp16 = tile(reps = k_exp_9_reps_0, x = var_616_cast_fp16)[name = string("k_exp_9_cast_fp16")];
            tensor<int32, [4]> var_619 = const()[name = string("op_619"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_11_cast_fp16 = reshape(shape = var_619, x = k_exp_9_cast_fp16)[name = string("k_exp_11_cast_fp16")];
            tensor<int32, [1]> var_621_axes_0 = const()[name = string("op_621_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_621_cast_fp16 = expand_dims(axes = var_621_axes_0, x = coreml_update_state_67)[name = string("op_621_cast_fp16")];
            tensor<int32, [5]> v_exp_9_reps_0 = const()[name = string("v_exp_9_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_9_cast_fp16 = tile(reps = v_exp_9_reps_0, x = var_621_cast_fp16)[name = string("v_exp_9_cast_fp16")];
            tensor<int32, [4]> var_624 = const()[name = string("op_624"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_11_cast_fp16 = reshape(shape = var_624, x = v_exp_9_cast_fp16)[name = string("v_exp_11_cast_fp16")];
            bool var_627_transpose_x_1 = const()[name = string("op_627_transpose_x_1"), val = bool(false)];
            bool var_627_transpose_y_1 = const()[name = string("op_627_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_627_cast_fp16 = matmul(transpose_x = var_627_transpose_x_1, transpose_y = var_627_transpose_y_1, x = q_5_cast_fp16, y = k_exp_11_cast_fp16)[name = string("op_627_cast_fp16")];
            fp16 var_628_to_fp16 = const()[name = string("op_628_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_9_cast_fp16 = mul(x = var_627_cast_fp16, y = var_628_to_fp16)[name = string("attn_9_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_21_cast_fp16 = add(x = attn_9_cast_fp16, y = attention_mask_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_11_cast_fp16 = softmax(axis = var_485, x = input_21_cast_fp16)[name = string("attn_11_cast_fp16")];
            bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)];
            bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_11_cast_fp16, y = v_exp_11_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<int32, [4]> var_633_perm_0 = const()[name = string("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_634 = const()[name = string("op_634"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = out_5_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 128, 2048]> input_23_cast_fp16 = reshape(shape = var_634, x = var_633_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35687680))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37784896))))[name = string("layers_2_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_71_cast_fp16 = add(x = x_51_cast_fp16, y = linear_17_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_484_promoted_3_to_fp16 = const()[name = string("op_484_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_641_cast_fp16 = pow(x = x_71_cast_fp16, y = var_484_promoted_3_to_fp16)[name = string("op_641_cast_fp16")];
            tensor<int32, [1]> var_643_axes_0 = const()[name = string("op_643_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_643_keep_dims_0 = const()[name = string("op_643_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_643_cast_fp16 = reduce_mean(axes = var_643_axes_0, keep_dims = var_643_keep_dims_0, x = var_641_cast_fp16)[name = string("op_643_cast_fp16")];
            fp16 var_644_to_fp16 = const()[name = string("op_644_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_645_cast_fp16 = add(x = var_643_cast_fp16, y = var_644_to_fp16)[name = string("op_645_cast_fp16")];
            fp32 norm_23_epsilon_0 = const()[name = string("norm_23_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_23_cast_fp16 = rsqrt(epsilon = norm_23_epsilon_0, x = var_645_cast_fp16)[name = string("norm_23_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_647_cast_fp16 = mul(x = x_71_cast_fp16, y = norm_23_cast_fp16)[name = string("op_647_cast_fp16")];
            tensor<fp16, [1024]> layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37785472)))];
            tensor<fp16, [1, 128, 1024]> var_648_cast_fp16 = mul(x = var_647_cast_fp16, y = layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_648_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_2_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37787584))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40933376))))[name = string("layers_2_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16_palettized, x = var_648_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_658_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_658_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_2_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40933952))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44079744))))[name = string("layers_2_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16_palettized, x = var_648_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_29_cast_fp16 = mul(x = var_658_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_2_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44080320))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47226112))))[name = string("layers_2_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_77_cast_fp16 = add(x = x_71_cast_fp16, y = linear_20_cast_fp16)[name = string("x_77_cast_fp16")];
            int32 var_679 = const()[name = string("op_679"), val = int32(-1)];
            fp16 var_678_promoted_to_fp16 = const()[name = string("op_678_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_688_cast_fp16 = pow(x = x_77_cast_fp16, y = var_678_promoted_to_fp16)[name = string("op_688_cast_fp16")];
            tensor<int32, [1]> var_690_axes_0 = const()[name = string("op_690_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_690_keep_dims_0 = const()[name = string("op_690_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_690_cast_fp16 = reduce_mean(axes = var_690_axes_0, keep_dims = var_690_keep_dims_0, x = var_688_cast_fp16)[name = string("op_690_cast_fp16")];
            fp16 var_691_to_fp16 = const()[name = string("op_691_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_692_cast_fp16 = add(x = var_690_cast_fp16, y = var_691_to_fp16)[name = string("op_692_cast_fp16")];
            fp32 norm_25_epsilon_0 = const()[name = string("norm_25_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_25_cast_fp16 = rsqrt(epsilon = norm_25_epsilon_0, x = var_692_cast_fp16)[name = string("norm_25_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_694_cast_fp16 = mul(x = x_77_cast_fp16, y = norm_25_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<fp16, [1024]> layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47226688)))];
            tensor<fp16, [1, 128, 1024]> var_695_cast_fp16 = mul(x = var_694_cast_fp16, y = layers_3_input_layernorm_weight_to_fp16)[name = string("op_695_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47228800))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49326016))))[name = string("layers_3_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = var_695_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_711 = const()[name = string("op_711"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_712_cast_fp16 = reshape(shape = var_711, x = linear_21_cast_fp16)[name = string("op_712_cast_fp16")];
            tensor<int32, [4]> x_83_perm_0 = const()[name = string("x_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49326592))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50375232))))[name = string("layers_3_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = var_695_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_716 = const()[name = string("op_716"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_717_cast_fp16 = reshape(shape = var_716, x = linear_22_cast_fp16)[name = string("op_717_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50375808))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51424448))))[name = string("layers_3_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = var_695_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_721 = const()[name = string("op_721"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_722_cast_fp16 = reshape(shape = var_721, x = linear_23_cast_fp16)[name = string("op_722_cast_fp16")];
            tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_678_promoted_1_to_fp16 = const()[name = string("op_678_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_83_cast_fp16 = transpose(perm = x_83_perm_0, x = var_712_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 16, 128, 128]> var_726_cast_fp16 = pow(x = x_83_cast_fp16, y = var_678_promoted_1_to_fp16)[name = string("op_726_cast_fp16")];
            tensor<int32, [1]> var_728_axes_0 = const()[name = string("op_728_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_728_keep_dims_0 = const()[name = string("op_728_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_728_cast_fp16 = reduce_mean(axes = var_728_axes_0, keep_dims = var_728_keep_dims_0, x = var_726_cast_fp16)[name = string("op_728_cast_fp16")];
            fp16 var_729_to_fp16 = const()[name = string("op_729_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_730_cast_fp16 = add(x = var_728_cast_fp16, y = var_729_to_fp16)[name = string("op_730_cast_fp16")];
            fp32 norm_27_epsilon_0 = const()[name = string("norm_27_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_27_cast_fp16 = rsqrt(epsilon = norm_27_epsilon_0, x = var_730_cast_fp16)[name = string("norm_27_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_732_cast_fp16 = mul(x = x_83_cast_fp16, y = norm_27_cast_fp16)[name = string("op_732_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425024)))];
            tensor<fp16, [1, 16, 128, 128]> var_733_cast_fp16 = mul(x = var_732_cast_fp16, y = layers_3_self_attn_q_norm_weight_to_fp16)[name = string("op_733_cast_fp16")];
            fp16 var_678_promoted_2_to_fp16 = const()[name = string("op_678_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_717_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 8, 128, 128]> var_737_cast_fp16 = pow(x = x_87_cast_fp16, y = var_678_promoted_2_to_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])];
            bool var_739_keep_dims_0 = const()[name = string("op_739_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_739_cast_fp16 = reduce_mean(axes = var_739_axes_0, keep_dims = var_739_keep_dims_0, x = var_737_cast_fp16)[name = string("op_739_cast_fp16")];
            fp16 var_740_to_fp16 = const()[name = string("op_740_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_741_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_to_fp16)[name = string("op_741_cast_fp16")];
            fp32 norm_29_epsilon_0 = const()[name = string("norm_29_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_29_cast_fp16 = rsqrt(epsilon = norm_29_epsilon_0, x = var_741_cast_fp16)[name = string("norm_29_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_743_cast_fp16 = mul(x = x_87_cast_fp16, y = norm_29_cast_fp16)[name = string("op_743_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425344)))];
            tensor<fp16, [1, 8, 128, 128]> var_744_cast_fp16 = mul(x = var_743_cast_fp16, y = layers_3_self_attn_k_norm_weight_to_fp16)[name = string("op_744_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_733_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, 64])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_733_cast_fp16)[name = string("x2_13_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_765_cast_fp16 = mul(x = x1_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_765_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_766_cast_fp16 = mul(x = x2_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_766_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_767_cast_fp16 = sub(x = var_765_cast_fp16, y = var_766_cast_fp16)[name = string("op_767_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_768_cast_fp16 = mul(x = x2_13_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_768_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_769_cast_fp16 = mul(x = x1_13_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_769_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_770_cast_fp16 = add(x = var_768_cast_fp16, y = var_769_cast_fp16)[name = string("op_770_cast_fp16")];
            bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_7_cast_fp16 = concat(axis = var_679, interleave = q_7_interleave_0, values = (var_767_cast_fp16, var_770_cast_fp16))[name = string("q_7_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_744_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, 64])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_744_cast_fp16)[name = string("x2_15_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_792_cast_fp16 = mul(x = x1_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_793_cast_fp16 = mul(x = x2_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_793_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_794_cast_fp16 = sub(x = var_792_cast_fp16, y = var_793_cast_fp16)[name = string("op_794_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_795_cast_fp16 = mul(x = x2_15_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_795_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_796_cast_fp16 = mul(x = x1_15_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_796_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_797_cast_fp16 = add(x = var_795_cast_fp16, y = var_796_cast_fp16)[name = string("op_797_cast_fp16")];
            bool var_799_interleave_0 = const()[name = string("op_799_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_799_cast_fp16 = concat(axis = var_679, interleave = var_799_interleave_0, values = (var_794_cast_fp16, var_797_cast_fp16))[name = string("op_799_cast_fp16")];
            tensor<int32, [4]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_58 = const()[name = string("concat_58"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = var_799_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [128, 1024]> reshape_19_cast_fp16 = reshape(shape = concat_58, x = transpose_13_cast_fp16)[name = string("reshape_19_cast_fp16")];
            bool matmul_6_transpose_x_1 = const()[name = string("matmul_6_transpose_x_1"), val = bool(true)];
            bool matmul_6_transpose_y_1 = const()[name = string("matmul_6_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_1, transpose_y = matmul_6_transpose_y_1, x = var_66_to_fp16, y = reshape_19_cast_fp16)[name = string("matmul_6_cast_fp16")];
            tensor<int32, [4]> concat_61 = const()[name = string("concat_61"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_20_cast_fp16 = reshape(shape = concat_61, x = matmul_6_cast_fp16)[name = string("reshape_20_cast_fp16")];
            tensor<int32, [4]> scattered_k_7_perm_0 = const()[name = string("scattered_k_7_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_66 = const()[name = string("concat_66"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_59_cast_fp16 = transpose(perm = transpose_59_perm_0, x = var_722_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [128, 1024]> reshape_22_cast_fp16 = reshape(shape = concat_66, x = transpose_59_cast_fp16)[name = string("reshape_22_cast_fp16")];
            bool matmul_7_transpose_x_1 = const()[name = string("matmul_7_transpose_x_1"), val = bool(true)];
            bool matmul_7_transpose_y_1 = const()[name = string("matmul_7_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_1, transpose_y = matmul_7_transpose_y_1, x = var_66_to_fp16, y = reshape_22_cast_fp16)[name = string("matmul_7_cast_fp16")];
            tensor<int32, [4]> concat_69 = const()[name = string("concat_69"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_23_cast_fp16 = reshape(shape = concat_69, x = matmul_7_cast_fp16)[name = string("reshape_23_cast_fp16")];
            tensor<int32, [4]> scattered_v_7_perm_0 = const()[name = string("scattered_v_7_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_6 = read_state(input = k_cache_3)[name = string("read_state_6")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_21_cast_fp16 = mul(x = read_state_6, y = var_222_cast_fp16)[name = string("k_cache_21_cast_fp16")];
            write_state(data = k_cache_21_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_68_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_68 = read_state(input = k_cache_3)[name = string("coreml_update_state_68")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_7_cast_fp16 = transpose(perm = scattered_k_7_perm_0, x = reshape_20_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_23_cast_fp16 = add(x = coreml_update_state_68, y = scattered_k_7_cast_fp16)[name = string("k_cache_23_cast_fp16")];
            write_state(data = k_cache_23_cast_fp16, input = k_cache_3)[name = string("coreml_update_state_69_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_69 = read_state(input = k_cache_3)[name = string("coreml_update_state_69")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_7 = read_state(input = v_cache_3)[name = string("read_state_7")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_21_cast_fp16 = mul(x = read_state_7, y = var_222_cast_fp16)[name = string("v_cache_21_cast_fp16")];
            write_state(data = v_cache_21_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_70_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_70 = read_state(input = v_cache_3)[name = string("coreml_update_state_70")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_7_cast_fp16 = transpose(perm = scattered_v_7_perm_0, x = reshape_23_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_23_cast_fp16 = add(x = coreml_update_state_70, y = scattered_v_7_cast_fp16)[name = string("v_cache_23_cast_fp16")];
            write_state(data = v_cache_23_cast_fp16, input = v_cache_3)[name = string("coreml_update_state_71_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_71 = read_state(input = v_cache_3)[name = string("coreml_update_state_71")];
            tensor<int32, [1]> var_810_axes_0 = const()[name = string("op_810_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_810_cast_fp16 = expand_dims(axes = var_810_axes_0, x = coreml_update_state_69)[name = string("op_810_cast_fp16")];
            tensor<int32, [5]> k_exp_13_reps_0 = const()[name = string("k_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_13_cast_fp16 = tile(reps = k_exp_13_reps_0, x = var_810_cast_fp16)[name = string("k_exp_13_cast_fp16")];
            tensor<int32, [4]> var_813 = const()[name = string("op_813"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_15_cast_fp16 = reshape(shape = var_813, x = k_exp_13_cast_fp16)[name = string("k_exp_15_cast_fp16")];
            tensor<int32, [1]> var_815_axes_0 = const()[name = string("op_815_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_815_cast_fp16 = expand_dims(axes = var_815_axes_0, x = coreml_update_state_71)[name = string("op_815_cast_fp16")];
            tensor<int32, [5]> v_exp_13_reps_0 = const()[name = string("v_exp_13_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_13_cast_fp16 = tile(reps = v_exp_13_reps_0, x = var_815_cast_fp16)[name = string("v_exp_13_cast_fp16")];
            tensor<int32, [4]> var_818 = const()[name = string("op_818"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_15_cast_fp16 = reshape(shape = var_818, x = v_exp_13_cast_fp16)[name = string("v_exp_15_cast_fp16")];
            bool var_821_transpose_x_1 = const()[name = string("op_821_transpose_x_1"), val = bool(false)];
            bool var_821_transpose_y_1 = const()[name = string("op_821_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_821_cast_fp16 = matmul(transpose_x = var_821_transpose_x_1, transpose_y = var_821_transpose_y_1, x = q_7_cast_fp16, y = k_exp_15_cast_fp16)[name = string("op_821_cast_fp16")];
            fp16 var_822_to_fp16 = const()[name = string("op_822_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_13_cast_fp16 = mul(x = var_821_cast_fp16, y = var_822_to_fp16)[name = string("attn_13_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_31_cast_fp16 = add(x = attn_13_cast_fp16, y = attention_mask_to_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_15_cast_fp16 = softmax(axis = var_679, x = input_31_cast_fp16)[name = string("attn_15_cast_fp16")];
            bool out_7_transpose_x_0 = const()[name = string("out_7_transpose_x_0"), val = bool(false)];
            bool out_7_transpose_y_0 = const()[name = string("out_7_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_15_cast_fp16, y = v_exp_15_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<int32, [4]> var_827_perm_0 = const()[name = string("op_827_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_828 = const()[name = string("op_828"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_827_cast_fp16 = transpose(perm = var_827_perm_0, x = out_7_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 128, 2048]> input_33_cast_fp16 = reshape(shape = var_828, x = var_827_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51425664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53522880))))[name = string("layers_3_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_97_cast_fp16 = add(x = x_77_cast_fp16, y = linear_24_cast_fp16)[name = string("x_97_cast_fp16")];
            fp16 var_678_promoted_3_to_fp16 = const()[name = string("op_678_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_835_cast_fp16 = pow(x = x_97_cast_fp16, y = var_678_promoted_3_to_fp16)[name = string("op_835_cast_fp16")];
            tensor<int32, [1]> var_837_axes_0 = const()[name = string("op_837_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_837_keep_dims_0 = const()[name = string("op_837_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_837_cast_fp16 = reduce_mean(axes = var_837_axes_0, keep_dims = var_837_keep_dims_0, x = var_835_cast_fp16)[name = string("op_837_cast_fp16")];
            fp16 var_838_to_fp16 = const()[name = string("op_838_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_839_cast_fp16 = add(x = var_837_cast_fp16, y = var_838_to_fp16)[name = string("op_839_cast_fp16")];
            fp32 norm_31_epsilon_0 = const()[name = string("norm_31_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_31_cast_fp16 = rsqrt(epsilon = norm_31_epsilon_0, x = var_839_cast_fp16)[name = string("norm_31_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_841_cast_fp16 = mul(x = x_97_cast_fp16, y = norm_31_cast_fp16)[name = string("op_841_cast_fp16")];
            tensor<fp16, [1024]> layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53523456)))];
            tensor<fp16, [1, 128, 1024]> var_842_cast_fp16 = mul(x = var_841_cast_fp16, y = layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_842_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_3_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53525568))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56671360))))[name = string("layers_3_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16_palettized, x = var_842_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_852_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_852_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_3_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56671936))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59817728))))[name = string("layers_3_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16_palettized, x = var_842_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_39_cast_fp16 = mul(x = var_852_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_3_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59818304))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62964096))))[name = string("layers_3_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16_palettized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_103_cast_fp16 = add(x = x_97_cast_fp16, y = linear_27_cast_fp16)[name = string("x_103_cast_fp16")];
            int32 var_873 = const()[name = string("op_873"), val = int32(-1)];
            fp16 var_872_promoted_to_fp16 = const()[name = string("op_872_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_882_cast_fp16 = pow(x = x_103_cast_fp16, y = var_872_promoted_to_fp16)[name = string("op_882_cast_fp16")];
            tensor<int32, [1]> var_884_axes_0 = const()[name = string("op_884_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_884_keep_dims_0 = const()[name = string("op_884_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_884_cast_fp16 = reduce_mean(axes = var_884_axes_0, keep_dims = var_884_keep_dims_0, x = var_882_cast_fp16)[name = string("op_884_cast_fp16")];
            fp16 var_885_to_fp16 = const()[name = string("op_885_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_886_cast_fp16 = add(x = var_884_cast_fp16, y = var_885_to_fp16)[name = string("op_886_cast_fp16")];
            fp32 norm_33_epsilon_0 = const()[name = string("norm_33_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_33_cast_fp16 = rsqrt(epsilon = norm_33_epsilon_0, x = var_886_cast_fp16)[name = string("norm_33_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_888_cast_fp16 = mul(x = x_103_cast_fp16, y = norm_33_cast_fp16)[name = string("op_888_cast_fp16")];
            tensor<fp16, [1024]> layers_4_input_layernorm_weight_to_fp16 = const()[name = string("layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62964672)))];
            tensor<fp16, [1, 128, 1024]> var_889_cast_fp16 = mul(x = var_888_cast_fp16, y = layers_4_input_layernorm_weight_to_fp16)[name = string("op_889_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62966784))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65064000))))[name = string("layers_4_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = var_889_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_905 = const()[name = string("op_905"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_906_cast_fp16 = reshape(shape = var_905, x = linear_28_cast_fp16)[name = string("op_906_cast_fp16")];
            tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65064576))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66113216))))[name = string("layers_4_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = var_889_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_910 = const()[name = string("op_910"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_911_cast_fp16 = reshape(shape = var_910, x = linear_29_cast_fp16)[name = string("op_911_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66113792))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67162432))))[name = string("layers_4_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = var_889_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_915 = const()[name = string("op_915"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_916_cast_fp16 = reshape(shape = var_915, x = linear_30_cast_fp16)[name = string("op_916_cast_fp16")];
            tensor<int32, [4]> transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_872_promoted_1_to_fp16 = const()[name = string("op_872_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_906_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 128, 128]> var_920_cast_fp16 = pow(x = x_109_cast_fp16, y = var_872_promoted_1_to_fp16)[name = string("op_920_cast_fp16")];
            tensor<int32, [1]> var_922_axes_0 = const()[name = string("op_922_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_922_keep_dims_0 = const()[name = string("op_922_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_922_cast_fp16 = reduce_mean(axes = var_922_axes_0, keep_dims = var_922_keep_dims_0, x = var_920_cast_fp16)[name = string("op_922_cast_fp16")];
            fp16 var_923_to_fp16 = const()[name = string("op_923_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_924_cast_fp16 = add(x = var_922_cast_fp16, y = var_923_to_fp16)[name = string("op_924_cast_fp16")];
            fp32 norm_35_epsilon_0 = const()[name = string("norm_35_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_35_cast_fp16 = rsqrt(epsilon = norm_35_epsilon_0, x = var_924_cast_fp16)[name = string("norm_35_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_926_cast_fp16 = mul(x = x_109_cast_fp16, y = norm_35_cast_fp16)[name = string("op_926_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163008)))];
            tensor<fp16, [1, 16, 128, 128]> var_927_cast_fp16 = mul(x = var_926_cast_fp16, y = layers_4_self_attn_q_norm_weight_to_fp16)[name = string("op_927_cast_fp16")];
            fp16 var_872_promoted_2_to_fp16 = const()[name = string("op_872_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_911_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 128, 128]> var_931_cast_fp16 = pow(x = x_113_cast_fp16, y = var_872_promoted_2_to_fp16)[name = string("op_931_cast_fp16")];
            tensor<int32, [1]> var_933_axes_0 = const()[name = string("op_933_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_933_keep_dims_0 = const()[name = string("op_933_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_933_cast_fp16 = reduce_mean(axes = var_933_axes_0, keep_dims = var_933_keep_dims_0, x = var_931_cast_fp16)[name = string("op_933_cast_fp16")];
            fp16 var_934_to_fp16 = const()[name = string("op_934_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_935_cast_fp16 = add(x = var_933_cast_fp16, y = var_934_to_fp16)[name = string("op_935_cast_fp16")];
            fp32 norm_37_epsilon_0 = const()[name = string("norm_37_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_37_cast_fp16 = rsqrt(epsilon = norm_37_epsilon_0, x = var_935_cast_fp16)[name = string("norm_37_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_937_cast_fp16 = mul(x = x_113_cast_fp16, y = norm_37_cast_fp16)[name = string("op_937_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163328)))];
            tensor<fp16, [1, 8, 128, 128]> var_938_cast_fp16 = mul(x = var_937_cast_fp16, y = layers_4_self_attn_k_norm_weight_to_fp16)[name = string("op_938_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_927_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, 64])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_927_cast_fp16)[name = string("x2_17_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_959_cast_fp16 = mul(x = x1_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_960_cast_fp16 = mul(x = x2_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_960_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_961_cast_fp16 = sub(x = var_959_cast_fp16, y = var_960_cast_fp16)[name = string("op_961_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_962_cast_fp16 = mul(x = x2_17_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_962_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_963_cast_fp16 = mul(x = x1_17_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_963_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_964_cast_fp16 = add(x = var_962_cast_fp16, y = var_963_cast_fp16)[name = string("op_964_cast_fp16")];
            bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_9_cast_fp16 = concat(axis = var_873, interleave = q_9_interleave_0, values = (var_961_cast_fp16, var_964_cast_fp16))[name = string("q_9_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_938_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, 64])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_938_cast_fp16)[name = string("x2_19_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_986_cast_fp16 = mul(x = x1_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_986_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_987_cast_fp16 = mul(x = x2_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_987_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_988_cast_fp16 = sub(x = var_986_cast_fp16, y = var_987_cast_fp16)[name = string("op_988_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_989_cast_fp16 = mul(x = x2_19_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_989_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_990_cast_fp16 = mul(x = x1_19_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_990_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_991_cast_fp16 = add(x = var_989_cast_fp16, y = var_990_cast_fp16)[name = string("op_991_cast_fp16")];
            bool var_993_interleave_0 = const()[name = string("op_993_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_993_cast_fp16 = concat(axis = var_873, interleave = var_993_interleave_0, values = (var_988_cast_fp16, var_991_cast_fp16))[name = string("op_993_cast_fp16")];
            tensor<int32, [4]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_76 = const()[name = string("concat_76"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = var_993_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [128, 1024]> reshape_25_cast_fp16 = reshape(shape = concat_76, x = transpose_17_cast_fp16)[name = string("reshape_25_cast_fp16")];
            bool matmul_8_transpose_x_1 = const()[name = string("matmul_8_transpose_x_1"), val = bool(true)];
            bool matmul_8_transpose_y_1 = const()[name = string("matmul_8_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_1, transpose_y = matmul_8_transpose_y_1, x = var_66_to_fp16, y = reshape_25_cast_fp16)[name = string("matmul_8_cast_fp16")];
            tensor<int32, [4]> concat_79 = const()[name = string("concat_79"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_26_cast_fp16 = reshape(shape = concat_79, x = matmul_8_cast_fp16)[name = string("reshape_26_cast_fp16")];
            tensor<int32, [4]> scattered_k_9_perm_0 = const()[name = string("scattered_k_9_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_84 = const()[name = string("concat_84"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_60_cast_fp16 = transpose(perm = transpose_60_perm_0, x = var_916_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [128, 1024]> reshape_28_cast_fp16 = reshape(shape = concat_84, x = transpose_60_cast_fp16)[name = string("reshape_28_cast_fp16")];
            bool matmul_9_transpose_x_1 = const()[name = string("matmul_9_transpose_x_1"), val = bool(true)];
            bool matmul_9_transpose_y_1 = const()[name = string("matmul_9_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_1, transpose_y = matmul_9_transpose_y_1, x = var_66_to_fp16, y = reshape_28_cast_fp16)[name = string("matmul_9_cast_fp16")];
            tensor<int32, [4]> concat_87 = const()[name = string("concat_87"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_29_cast_fp16 = reshape(shape = concat_87, x = matmul_9_cast_fp16)[name = string("reshape_29_cast_fp16")];
            tensor<int32, [4]> scattered_v_9_perm_0 = const()[name = string("scattered_v_9_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_8 = read_state(input = k_cache_4)[name = string("read_state_8")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_27_cast_fp16 = mul(x = read_state_8, y = var_222_cast_fp16)[name = string("k_cache_27_cast_fp16")];
            write_state(data = k_cache_27_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_72_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_72 = read_state(input = k_cache_4)[name = string("coreml_update_state_72")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_9_cast_fp16 = transpose(perm = scattered_k_9_perm_0, x = reshape_26_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_29_cast_fp16 = add(x = coreml_update_state_72, y = scattered_k_9_cast_fp16)[name = string("k_cache_29_cast_fp16")];
            write_state(data = k_cache_29_cast_fp16, input = k_cache_4)[name = string("coreml_update_state_73_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_73 = read_state(input = k_cache_4)[name = string("coreml_update_state_73")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_9 = read_state(input = v_cache_4)[name = string("read_state_9")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_27_cast_fp16 = mul(x = read_state_9, y = var_222_cast_fp16)[name = string("v_cache_27_cast_fp16")];
            write_state(data = v_cache_27_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_74_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_74 = read_state(input = v_cache_4)[name = string("coreml_update_state_74")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_9_cast_fp16 = transpose(perm = scattered_v_9_perm_0, x = reshape_29_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_29_cast_fp16 = add(x = coreml_update_state_74, y = scattered_v_9_cast_fp16)[name = string("v_cache_29_cast_fp16")];
            write_state(data = v_cache_29_cast_fp16, input = v_cache_4)[name = string("coreml_update_state_75_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_75 = read_state(input = v_cache_4)[name = string("coreml_update_state_75")];
            tensor<int32, [1]> var_1004_axes_0 = const()[name = string("op_1004_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1004_cast_fp16 = expand_dims(axes = var_1004_axes_0, x = coreml_update_state_73)[name = string("op_1004_cast_fp16")];
            tensor<int32, [5]> k_exp_17_reps_0 = const()[name = string("k_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_17_cast_fp16 = tile(reps = k_exp_17_reps_0, x = var_1004_cast_fp16)[name = string("k_exp_17_cast_fp16")];
            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_19_cast_fp16 = reshape(shape = var_1007, x = k_exp_17_cast_fp16)[name = string("k_exp_19_cast_fp16")];
            tensor<int32, [1]> var_1009_axes_0 = const()[name = string("op_1009_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1009_cast_fp16 = expand_dims(axes = var_1009_axes_0, x = coreml_update_state_75)[name = string("op_1009_cast_fp16")];
            tensor<int32, [5]> v_exp_17_reps_0 = const()[name = string("v_exp_17_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_17_cast_fp16 = tile(reps = v_exp_17_reps_0, x = var_1009_cast_fp16)[name = string("v_exp_17_cast_fp16")];
            tensor<int32, [4]> var_1012 = const()[name = string("op_1012"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_19_cast_fp16 = reshape(shape = var_1012, x = v_exp_17_cast_fp16)[name = string("v_exp_19_cast_fp16")];
            bool var_1015_transpose_x_1 = const()[name = string("op_1015_transpose_x_1"), val = bool(false)];
            bool var_1015_transpose_y_1 = const()[name = string("op_1015_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1015_cast_fp16 = matmul(transpose_x = var_1015_transpose_x_1, transpose_y = var_1015_transpose_y_1, x = q_9_cast_fp16, y = k_exp_19_cast_fp16)[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_17_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("attn_17_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_41_cast_fp16 = add(x = attn_17_cast_fp16, y = attention_mask_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_19_cast_fp16 = softmax(axis = var_873, x = input_41_cast_fp16)[name = string("attn_19_cast_fp16")];
            bool out_9_transpose_x_0 = const()[name = string("out_9_transpose_x_0"), val = bool(false)];
            bool out_9_transpose_y_0 = const()[name = string("out_9_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_19_cast_fp16, y = v_exp_19_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<int32, [4]> var_1021_perm_0 = const()[name = string("op_1021_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1022 = const()[name = string("op_1022"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1021_cast_fp16 = transpose(perm = var_1021_perm_0, x = out_9_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 128, 2048]> input_43_cast_fp16 = reshape(shape = var_1022, x = var_1021_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_4_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67163648))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69260864))))[name = string("layers_4_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_123_cast_fp16 = add(x = x_103_cast_fp16, y = linear_31_cast_fp16)[name = string("x_123_cast_fp16")];
            fp16 var_872_promoted_3_to_fp16 = const()[name = string("op_872_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1029_cast_fp16 = pow(x = x_123_cast_fp16, y = var_872_promoted_3_to_fp16)[name = string("op_1029_cast_fp16")];
            tensor<int32, [1]> var_1031_axes_0 = const()[name = string("op_1031_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1031_keep_dims_0 = const()[name = string("op_1031_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1031_cast_fp16 = reduce_mean(axes = var_1031_axes_0, keep_dims = var_1031_keep_dims_0, x = var_1029_cast_fp16)[name = string("op_1031_cast_fp16")];
            fp16 var_1032_to_fp16 = const()[name = string("op_1032_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1033_cast_fp16 = add(x = var_1031_cast_fp16, y = var_1032_to_fp16)[name = string("op_1033_cast_fp16")];
            fp32 norm_39_epsilon_0 = const()[name = string("norm_39_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_39_cast_fp16 = rsqrt(epsilon = norm_39_epsilon_0, x = var_1033_cast_fp16)[name = string("norm_39_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1035_cast_fp16 = mul(x = x_123_cast_fp16, y = norm_39_cast_fp16)[name = string("op_1035_cast_fp16")];
            tensor<fp16, [1024]> layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69261440)))];
            tensor<fp16, [1, 128, 1024]> var_1036_cast_fp16 = mul(x = var_1035_cast_fp16, y = layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_1036_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_4_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69263552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72409344))))[name = string("layers_4_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16_palettized, x = var_1036_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_1046_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_1046_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_4_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72409920))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75555712))))[name = string("layers_4_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16_palettized, x = var_1036_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_49_cast_fp16 = mul(x = var_1046_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_4_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75556288))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78702080))))[name = string("layers_4_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_129_cast_fp16 = add(x = x_123_cast_fp16, y = linear_34_cast_fp16)[name = string("x_129_cast_fp16")];
            int32 var_1067 = const()[name = string("op_1067"), val = int32(-1)];
            fp16 var_1066_promoted_to_fp16 = const()[name = string("op_1066_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1076_cast_fp16 = pow(x = x_129_cast_fp16, y = var_1066_promoted_to_fp16)[name = string("op_1076_cast_fp16")];
            tensor<int32, [1]> var_1078_axes_0 = const()[name = string("op_1078_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1078_keep_dims_0 = const()[name = string("op_1078_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1078_cast_fp16 = reduce_mean(axes = var_1078_axes_0, keep_dims = var_1078_keep_dims_0, x = var_1076_cast_fp16)[name = string("op_1078_cast_fp16")];
            fp16 var_1079_to_fp16 = const()[name = string("op_1079_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1080_cast_fp16 = add(x = var_1078_cast_fp16, y = var_1079_to_fp16)[name = string("op_1080_cast_fp16")];
            fp32 norm_41_epsilon_0 = const()[name = string("norm_41_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_41_cast_fp16 = rsqrt(epsilon = norm_41_epsilon_0, x = var_1080_cast_fp16)[name = string("norm_41_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1082_cast_fp16 = mul(x = x_129_cast_fp16, y = norm_41_cast_fp16)[name = string("op_1082_cast_fp16")];
            tensor<fp16, [1024]> layers_5_input_layernorm_weight_to_fp16 = const()[name = string("layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78702656)))];
            tensor<fp16, [1, 128, 1024]> var_1083_cast_fp16 = mul(x = var_1082_cast_fp16, y = layers_5_input_layernorm_weight_to_fp16)[name = string("op_1083_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78704768))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80801984))))[name = string("layers_5_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = var_1083_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_1099 = const()[name = string("op_1099"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_1100_cast_fp16 = reshape(shape = var_1099, x = linear_35_cast_fp16)[name = string("op_1100_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80802560))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81851200))))[name = string("layers_5_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = var_1083_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_1104 = const()[name = string("op_1104"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1105_cast_fp16 = reshape(shape = var_1104, x = linear_36_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<int32, [4]> x_139_perm_0 = const()[name = string("x_139_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81851776))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82900416))))[name = string("layers_5_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = var_1083_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = string("op_1109"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1110_cast_fp16 = reshape(shape = var_1109, x = linear_37_cast_fp16)[name = string("op_1110_cast_fp16")];
            tensor<int32, [4]> transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_1066_promoted_1_to_fp16 = const()[name = string("op_1066_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_1100_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 16, 128, 128]> var_1114_cast_fp16 = pow(x = x_135_cast_fp16, y = var_1066_promoted_1_to_fp16)[name = string("op_1114_cast_fp16")];
            tensor<int32, [1]> var_1116_axes_0 = const()[name = string("op_1116_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1116_keep_dims_0 = const()[name = string("op_1116_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_1116_cast_fp16 = reduce_mean(axes = var_1116_axes_0, keep_dims = var_1116_keep_dims_0, x = var_1114_cast_fp16)[name = string("op_1116_cast_fp16")];
            fp16 var_1117_to_fp16 = const()[name = string("op_1117_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_1118_cast_fp16 = add(x = var_1116_cast_fp16, y = var_1117_to_fp16)[name = string("op_1118_cast_fp16")];
            fp32 norm_43_epsilon_0 = const()[name = string("norm_43_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_43_cast_fp16 = rsqrt(epsilon = norm_43_epsilon_0, x = var_1118_cast_fp16)[name = string("norm_43_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1120_cast_fp16 = mul(x = x_135_cast_fp16, y = norm_43_cast_fp16)[name = string("op_1120_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82900992)))];
            tensor<fp16, [1, 16, 128, 128]> var_1121_cast_fp16 = mul(x = var_1120_cast_fp16, y = layers_5_self_attn_q_norm_weight_to_fp16)[name = string("op_1121_cast_fp16")];
            fp16 var_1066_promoted_2_to_fp16 = const()[name = string("op_1066_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_139_cast_fp16 = transpose(perm = x_139_perm_0, x = var_1105_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 128, 128]> var_1125_cast_fp16 = pow(x = x_139_cast_fp16, y = var_1066_promoted_2_to_fp16)[name = string("op_1125_cast_fp16")];
            tensor<int32, [1]> var_1127_axes_0 = const()[name = string("op_1127_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1127_keep_dims_0 = const()[name = string("op_1127_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_1127_cast_fp16 = reduce_mean(axes = var_1127_axes_0, keep_dims = var_1127_keep_dims_0, x = var_1125_cast_fp16)[name = string("op_1127_cast_fp16")];
            fp16 var_1128_to_fp16 = const()[name = string("op_1128_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_1129_cast_fp16 = add(x = var_1127_cast_fp16, y = var_1128_to_fp16)[name = string("op_1129_cast_fp16")];
            fp32 norm_45_epsilon_0 = const()[name = string("norm_45_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_45_cast_fp16 = rsqrt(epsilon = norm_45_epsilon_0, x = var_1129_cast_fp16)[name = string("norm_45_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_1131_cast_fp16 = mul(x = x_139_cast_fp16, y = norm_45_cast_fp16)[name = string("op_1131_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82901312)))];
            tensor<fp16, [1, 8, 128, 128]> var_1132_cast_fp16 = mul(x = var_1131_cast_fp16, y = layers_5_self_attn_k_norm_weight_to_fp16)[name = string("op_1132_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_1121_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, 64])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_1121_cast_fp16)[name = string("x2_21_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1153_cast_fp16 = mul(x = x1_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1153_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1154_cast_fp16 = mul(x = x2_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1154_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1155_cast_fp16 = sub(x = var_1153_cast_fp16, y = var_1154_cast_fp16)[name = string("op_1155_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1156_cast_fp16 = mul(x = x2_21_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1156_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1157_cast_fp16 = mul(x = x1_21_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1157_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1158_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1157_cast_fp16)[name = string("op_1158_cast_fp16")];
            bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_11_cast_fp16 = concat(axis = var_1067, interleave = q_11_interleave_0, values = (var_1155_cast_fp16, var_1158_cast_fp16))[name = string("q_11_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_1132_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, 64])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_1132_cast_fp16)[name = string("x2_23_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1180_cast_fp16 = mul(x = x1_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1180_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1181_cast_fp16 = mul(x = x2_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1181_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1182_cast_fp16 = sub(x = var_1180_cast_fp16, y = var_1181_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1183_cast_fp16 = mul(x = x2_23_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1183_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1184_cast_fp16 = mul(x = x1_23_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1184_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1185_cast_fp16 = add(x = var_1183_cast_fp16, y = var_1184_cast_fp16)[name = string("op_1185_cast_fp16")];
            bool var_1187_interleave_0 = const()[name = string("op_1187_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1187_cast_fp16 = concat(axis = var_1067, interleave = var_1187_interleave_0, values = (var_1182_cast_fp16, var_1185_cast_fp16))[name = string("op_1187_cast_fp16")];
            tensor<int32, [4]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_94 = const()[name = string("concat_94"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = var_1187_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [128, 1024]> reshape_31_cast_fp16 = reshape(shape = concat_94, x = transpose_21_cast_fp16)[name = string("reshape_31_cast_fp16")];
            bool matmul_10_transpose_x_1 = const()[name = string("matmul_10_transpose_x_1"), val = bool(true)];
            bool matmul_10_transpose_y_1 = const()[name = string("matmul_10_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_1, transpose_y = matmul_10_transpose_y_1, x = var_66_to_fp16, y = reshape_31_cast_fp16)[name = string("matmul_10_cast_fp16")];
            tensor<int32, [4]> concat_97 = const()[name = string("concat_97"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_32_cast_fp16 = reshape(shape = concat_97, x = matmul_10_cast_fp16)[name = string("reshape_32_cast_fp16")];
            tensor<int32, [4]> scattered_k_11_perm_0 = const()[name = string("scattered_k_11_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_102 = const()[name = string("concat_102"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_61_cast_fp16 = transpose(perm = transpose_61_perm_0, x = var_1110_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [128, 1024]> reshape_34_cast_fp16 = reshape(shape = concat_102, x = transpose_61_cast_fp16)[name = string("reshape_34_cast_fp16")];
            bool matmul_11_transpose_x_1 = const()[name = string("matmul_11_transpose_x_1"), val = bool(true)];
            bool matmul_11_transpose_y_1 = const()[name = string("matmul_11_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_1, transpose_y = matmul_11_transpose_y_1, x = var_66_to_fp16, y = reshape_34_cast_fp16)[name = string("matmul_11_cast_fp16")];
            tensor<int32, [4]> concat_105 = const()[name = string("concat_105"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_35_cast_fp16 = reshape(shape = concat_105, x = matmul_11_cast_fp16)[name = string("reshape_35_cast_fp16")];
            tensor<int32, [4]> scattered_v_11_perm_0 = const()[name = string("scattered_v_11_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_10 = read_state(input = k_cache_5)[name = string("read_state_10")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_33_cast_fp16 = mul(x = read_state_10, y = var_222_cast_fp16)[name = string("k_cache_33_cast_fp16")];
            write_state(data = k_cache_33_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_76_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_76 = read_state(input = k_cache_5)[name = string("coreml_update_state_76")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_11_cast_fp16 = transpose(perm = scattered_k_11_perm_0, x = reshape_32_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_35_cast_fp16 = add(x = coreml_update_state_76, y = scattered_k_11_cast_fp16)[name = string("k_cache_35_cast_fp16")];
            write_state(data = k_cache_35_cast_fp16, input = k_cache_5)[name = string("coreml_update_state_77_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_77 = read_state(input = k_cache_5)[name = string("coreml_update_state_77")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_11 = read_state(input = v_cache_5)[name = string("read_state_11")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_33_cast_fp16 = mul(x = read_state_11, y = var_222_cast_fp16)[name = string("v_cache_33_cast_fp16")];
            write_state(data = v_cache_33_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_78_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_78 = read_state(input = v_cache_5)[name = string("coreml_update_state_78")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_11_cast_fp16 = transpose(perm = scattered_v_11_perm_0, x = reshape_35_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_35_cast_fp16 = add(x = coreml_update_state_78, y = scattered_v_11_cast_fp16)[name = string("v_cache_35_cast_fp16")];
            write_state(data = v_cache_35_cast_fp16, input = v_cache_5)[name = string("coreml_update_state_79_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_79 = read_state(input = v_cache_5)[name = string("coreml_update_state_79")];
            tensor<int32, [1]> var_1198_axes_0 = const()[name = string("op_1198_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1198_cast_fp16 = expand_dims(axes = var_1198_axes_0, x = coreml_update_state_77)[name = string("op_1198_cast_fp16")];
            tensor<int32, [5]> k_exp_21_reps_0 = const()[name = string("k_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_21_cast_fp16 = tile(reps = k_exp_21_reps_0, x = var_1198_cast_fp16)[name = string("k_exp_21_cast_fp16")];
            tensor<int32, [4]> var_1201 = const()[name = string("op_1201"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_23_cast_fp16 = reshape(shape = var_1201, x = k_exp_21_cast_fp16)[name = string("k_exp_23_cast_fp16")];
            tensor<int32, [1]> var_1203_axes_0 = const()[name = string("op_1203_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1203_cast_fp16 = expand_dims(axes = var_1203_axes_0, x = coreml_update_state_79)[name = string("op_1203_cast_fp16")];
            tensor<int32, [5]> v_exp_21_reps_0 = const()[name = string("v_exp_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_21_cast_fp16 = tile(reps = v_exp_21_reps_0, x = var_1203_cast_fp16)[name = string("v_exp_21_cast_fp16")];
            tensor<int32, [4]> var_1206 = const()[name = string("op_1206"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_23_cast_fp16 = reshape(shape = var_1206, x = v_exp_21_cast_fp16)[name = string("v_exp_23_cast_fp16")];
            bool var_1209_transpose_x_1 = const()[name = string("op_1209_transpose_x_1"), val = bool(false)];
            bool var_1209_transpose_y_1 = const()[name = string("op_1209_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1209_cast_fp16 = matmul(transpose_x = var_1209_transpose_x_1, transpose_y = var_1209_transpose_y_1, x = q_11_cast_fp16, y = k_exp_23_cast_fp16)[name = string("op_1209_cast_fp16")];
            fp16 var_1210_to_fp16 = const()[name = string("op_1210_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_21_cast_fp16 = mul(x = var_1209_cast_fp16, y = var_1210_to_fp16)[name = string("attn_21_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_51_cast_fp16 = add(x = attn_21_cast_fp16, y = attention_mask_to_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_23_cast_fp16 = softmax(axis = var_1067, x = input_51_cast_fp16)[name = string("attn_23_cast_fp16")];
            bool out_11_transpose_x_0 = const()[name = string("out_11_transpose_x_0"), val = bool(false)];
            bool out_11_transpose_y_0 = const()[name = string("out_11_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_23_cast_fp16, y = v_exp_23_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<int32, [4]> var_1215_perm_0 = const()[name = string("op_1215_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1216 = const()[name = string("op_1216"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1215_cast_fp16 = transpose(perm = var_1215_perm_0, x = out_11_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 128, 2048]> input_53_cast_fp16 = reshape(shape = var_1216, x = var_1215_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_5_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82901632))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84998848))))[name = string("layers_5_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16_palettized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_149_cast_fp16 = add(x = x_129_cast_fp16, y = linear_38_cast_fp16)[name = string("x_149_cast_fp16")];
            fp16 var_1066_promoted_3_to_fp16 = const()[name = string("op_1066_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1223_cast_fp16 = pow(x = x_149_cast_fp16, y = var_1066_promoted_3_to_fp16)[name = string("op_1223_cast_fp16")];
            tensor<int32, [1]> var_1225_axes_0 = const()[name = string("op_1225_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1225_keep_dims_0 = const()[name = string("op_1225_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1225_cast_fp16 = reduce_mean(axes = var_1225_axes_0, keep_dims = var_1225_keep_dims_0, x = var_1223_cast_fp16)[name = string("op_1225_cast_fp16")];
            fp16 var_1226_to_fp16 = const()[name = string("op_1226_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1227_cast_fp16 = add(x = var_1225_cast_fp16, y = var_1226_to_fp16)[name = string("op_1227_cast_fp16")];
            fp32 norm_47_epsilon_0 = const()[name = string("norm_47_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_47_cast_fp16 = rsqrt(epsilon = norm_47_epsilon_0, x = var_1227_cast_fp16)[name = string("norm_47_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1229_cast_fp16 = mul(x = x_149_cast_fp16, y = norm_47_cast_fp16)[name = string("op_1229_cast_fp16")];
            tensor<fp16, [1024]> layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84999424)))];
            tensor<fp16, [1, 128, 1024]> var_1230_cast_fp16 = mul(x = var_1229_cast_fp16, y = layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_1230_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_5_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85001536))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88147328))))[name = string("layers_5_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16_palettized, x = var_1230_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_1240_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_1240_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_5_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88147904))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91293696))))[name = string("layers_5_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16_palettized, x = var_1230_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_59_cast_fp16 = mul(x = var_1240_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_5_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91294272))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94440064))))[name = string("layers_5_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16_palettized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_155_cast_fp16 = add(x = x_149_cast_fp16, y = linear_41_cast_fp16)[name = string("x_155_cast_fp16")];
            int32 var_1261 = const()[name = string("op_1261"), val = int32(-1)];
            fp16 var_1260_promoted_to_fp16 = const()[name = string("op_1260_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1270_cast_fp16 = pow(x = x_155_cast_fp16, y = var_1260_promoted_to_fp16)[name = string("op_1270_cast_fp16")];
            tensor<int32, [1]> var_1272_axes_0 = const()[name = string("op_1272_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1272_keep_dims_0 = const()[name = string("op_1272_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1272_cast_fp16 = reduce_mean(axes = var_1272_axes_0, keep_dims = var_1272_keep_dims_0, x = var_1270_cast_fp16)[name = string("op_1272_cast_fp16")];
            fp16 var_1273_to_fp16 = const()[name = string("op_1273_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1274_cast_fp16 = add(x = var_1272_cast_fp16, y = var_1273_to_fp16)[name = string("op_1274_cast_fp16")];
            fp32 norm_49_epsilon_0 = const()[name = string("norm_49_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_49_cast_fp16 = rsqrt(epsilon = norm_49_epsilon_0, x = var_1274_cast_fp16)[name = string("norm_49_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1276_cast_fp16 = mul(x = x_155_cast_fp16, y = norm_49_cast_fp16)[name = string("op_1276_cast_fp16")];
            tensor<fp16, [1024]> layers_6_input_layernorm_weight_to_fp16 = const()[name = string("layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94440640)))];
            tensor<fp16, [1, 128, 1024]> var_1277_cast_fp16 = mul(x = var_1276_cast_fp16, y = layers_6_input_layernorm_weight_to_fp16)[name = string("op_1277_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94442752))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96539968))))[name = string("layers_6_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = var_1277_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1293 = const()[name = string("op_1293"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_1294_cast_fp16 = reshape(shape = var_1293, x = linear_42_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96540544))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589184))))[name = string("layers_6_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = var_1277_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1298 = const()[name = string("op_1298"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1299_cast_fp16 = reshape(shape = var_1298, x = linear_43_cast_fp16)[name = string("op_1299_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589760))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98638400))))[name = string("layers_6_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = var_1277_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1303 = const()[name = string("op_1303"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1304_cast_fp16 = reshape(shape = var_1303, x = linear_44_cast_fp16)[name = string("op_1304_cast_fp16")];
            tensor<int32, [4]> transpose_62_perm_0 = const()[name = string("transpose_62_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_1260_promoted_1_to_fp16 = const()[name = string("op_1260_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_1294_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 128, 128]> var_1308_cast_fp16 = pow(x = x_161_cast_fp16, y = var_1260_promoted_1_to_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])];
            bool var_1310_keep_dims_0 = const()[name = string("op_1310_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_1310_cast_fp16 = reduce_mean(axes = var_1310_axes_0, keep_dims = var_1310_keep_dims_0, x = var_1308_cast_fp16)[name = string("op_1310_cast_fp16")];
            fp16 var_1311_to_fp16 = const()[name = string("op_1311_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_1312_cast_fp16 = add(x = var_1310_cast_fp16, y = var_1311_to_fp16)[name = string("op_1312_cast_fp16")];
            fp32 norm_51_epsilon_0 = const()[name = string("norm_51_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_51_cast_fp16 = rsqrt(epsilon = norm_51_epsilon_0, x = var_1312_cast_fp16)[name = string("norm_51_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1314_cast_fp16 = mul(x = x_161_cast_fp16, y = norm_51_cast_fp16)[name = string("op_1314_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98638976)))];
            tensor<fp16, [1, 16, 128, 128]> var_1315_cast_fp16 = mul(x = var_1314_cast_fp16, y = layers_6_self_attn_q_norm_weight_to_fp16)[name = string("op_1315_cast_fp16")];
            fp16 var_1260_promoted_2_to_fp16 = const()[name = string("op_1260_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_1299_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 8, 128, 128]> var_1319_cast_fp16 = pow(x = x_165_cast_fp16, y = var_1260_promoted_2_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<int32, [1]> var_1321_axes_0 = const()[name = string("op_1321_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1321_keep_dims_0 = const()[name = string("op_1321_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_1321_cast_fp16 = reduce_mean(axes = var_1321_axes_0, keep_dims = var_1321_keep_dims_0, x = var_1319_cast_fp16)[name = string("op_1321_cast_fp16")];
            fp16 var_1322_to_fp16 = const()[name = string("op_1322_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_1323_cast_fp16 = add(x = var_1321_cast_fp16, y = var_1322_to_fp16)[name = string("op_1323_cast_fp16")];
            fp32 norm_53_epsilon_0 = const()[name = string("norm_53_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_53_cast_fp16 = rsqrt(epsilon = norm_53_epsilon_0, x = var_1323_cast_fp16)[name = string("norm_53_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_1325_cast_fp16 = mul(x = x_165_cast_fp16, y = norm_53_cast_fp16)[name = string("op_1325_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98639296)))];
            tensor<fp16, [1, 8, 128, 128]> var_1326_cast_fp16 = mul(x = var_1325_cast_fp16, y = layers_6_self_attn_k_norm_weight_to_fp16)[name = string("op_1326_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_1315_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, 64])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_1315_cast_fp16)[name = string("x2_25_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1347_cast_fp16 = mul(x = x1_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1347_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1348_cast_fp16 = mul(x = x2_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1348_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1349_cast_fp16 = sub(x = var_1347_cast_fp16, y = var_1348_cast_fp16)[name = string("op_1349_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1350_cast_fp16 = mul(x = x2_25_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1350_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1351_cast_fp16 = mul(x = x1_25_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1351_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1352_cast_fp16 = add(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = string("op_1352_cast_fp16")];
            bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_13_cast_fp16 = concat(axis = var_1261, interleave = q_13_interleave_0, values = (var_1349_cast_fp16, var_1352_cast_fp16))[name = string("q_13_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_1326_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, 64])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_1326_cast_fp16)[name = string("x2_27_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1374_cast_fp16 = mul(x = x1_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1374_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1375_cast_fp16 = mul(x = x2_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1375_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1376_cast_fp16 = sub(x = var_1374_cast_fp16, y = var_1375_cast_fp16)[name = string("op_1376_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1377_cast_fp16 = mul(x = x2_27_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1377_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1378_cast_fp16 = mul(x = x1_27_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1378_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1379_cast_fp16 = add(x = var_1377_cast_fp16, y = var_1378_cast_fp16)[name = string("op_1379_cast_fp16")];
            bool var_1381_interleave_0 = const()[name = string("op_1381_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1381_cast_fp16 = concat(axis = var_1261, interleave = var_1381_interleave_0, values = (var_1376_cast_fp16, var_1379_cast_fp16))[name = string("op_1381_cast_fp16")];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_112 = const()[name = string("concat_112"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = var_1381_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [128, 1024]> reshape_37_cast_fp16 = reshape(shape = concat_112, x = transpose_25_cast_fp16)[name = string("reshape_37_cast_fp16")];
            bool matmul_12_transpose_x_1 = const()[name = string("matmul_12_transpose_x_1"), val = bool(true)];
            bool matmul_12_transpose_y_1 = const()[name = string("matmul_12_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_12_cast_fp16 = matmul(transpose_x = matmul_12_transpose_x_1, transpose_y = matmul_12_transpose_y_1, x = var_66_to_fp16, y = reshape_37_cast_fp16)[name = string("matmul_12_cast_fp16")];
            tensor<int32, [4]> concat_115 = const()[name = string("concat_115"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_38_cast_fp16 = reshape(shape = concat_115, x = matmul_12_cast_fp16)[name = string("reshape_38_cast_fp16")];
            tensor<int32, [4]> scattered_k_13_perm_0 = const()[name = string("scattered_k_13_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_120 = const()[name = string("concat_120"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_62_cast_fp16 = transpose(perm = transpose_62_perm_0, x = var_1304_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [128, 1024]> reshape_40_cast_fp16 = reshape(shape = concat_120, x = transpose_62_cast_fp16)[name = string("reshape_40_cast_fp16")];
            bool matmul_13_transpose_x_1 = const()[name = string("matmul_13_transpose_x_1"), val = bool(true)];
            bool matmul_13_transpose_y_1 = const()[name = string("matmul_13_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_1, transpose_y = matmul_13_transpose_y_1, x = var_66_to_fp16, y = reshape_40_cast_fp16)[name = string("matmul_13_cast_fp16")];
            tensor<int32, [4]> concat_123 = const()[name = string("concat_123"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_41_cast_fp16 = reshape(shape = concat_123, x = matmul_13_cast_fp16)[name = string("reshape_41_cast_fp16")];
            tensor<int32, [4]> scattered_v_13_perm_0 = const()[name = string("scattered_v_13_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_12 = read_state(input = k_cache_6)[name = string("read_state_12")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_39_cast_fp16 = mul(x = read_state_12, y = var_222_cast_fp16)[name = string("k_cache_39_cast_fp16")];
            write_state(data = k_cache_39_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_80_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_80 = read_state(input = k_cache_6)[name = string("coreml_update_state_80")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_13_cast_fp16 = transpose(perm = scattered_k_13_perm_0, x = reshape_38_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_41_cast_fp16 = add(x = coreml_update_state_80, y = scattered_k_13_cast_fp16)[name = string("k_cache_41_cast_fp16")];
            write_state(data = k_cache_41_cast_fp16, input = k_cache_6)[name = string("coreml_update_state_81_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_81 = read_state(input = k_cache_6)[name = string("coreml_update_state_81")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_13 = read_state(input = v_cache_6)[name = string("read_state_13")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_39_cast_fp16 = mul(x = read_state_13, y = var_222_cast_fp16)[name = string("v_cache_39_cast_fp16")];
            write_state(data = v_cache_39_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_82_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_82 = read_state(input = v_cache_6)[name = string("coreml_update_state_82")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_13_cast_fp16 = transpose(perm = scattered_v_13_perm_0, x = reshape_41_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_41_cast_fp16 = add(x = coreml_update_state_82, y = scattered_v_13_cast_fp16)[name = string("v_cache_41_cast_fp16")];
            write_state(data = v_cache_41_cast_fp16, input = v_cache_6)[name = string("coreml_update_state_83_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_83 = read_state(input = v_cache_6)[name = string("coreml_update_state_83")];
            tensor<int32, [1]> var_1392_axes_0 = const()[name = string("op_1392_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1392_cast_fp16 = expand_dims(axes = var_1392_axes_0, x = coreml_update_state_81)[name = string("op_1392_cast_fp16")];
            tensor<int32, [5]> k_exp_25_reps_0 = const()[name = string("k_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_25_cast_fp16 = tile(reps = k_exp_25_reps_0, x = var_1392_cast_fp16)[name = string("k_exp_25_cast_fp16")];
            tensor<int32, [4]> var_1395 = const()[name = string("op_1395"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_27_cast_fp16 = reshape(shape = var_1395, x = k_exp_25_cast_fp16)[name = string("k_exp_27_cast_fp16")];
            tensor<int32, [1]> var_1397_axes_0 = const()[name = string("op_1397_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1397_cast_fp16 = expand_dims(axes = var_1397_axes_0, x = coreml_update_state_83)[name = string("op_1397_cast_fp16")];
            tensor<int32, [5]> v_exp_25_reps_0 = const()[name = string("v_exp_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_25_cast_fp16 = tile(reps = v_exp_25_reps_0, x = var_1397_cast_fp16)[name = string("v_exp_25_cast_fp16")];
            tensor<int32, [4]> var_1400 = const()[name = string("op_1400"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_27_cast_fp16 = reshape(shape = var_1400, x = v_exp_25_cast_fp16)[name = string("v_exp_27_cast_fp16")];
            bool var_1403_transpose_x_1 = const()[name = string("op_1403_transpose_x_1"), val = bool(false)];
            bool var_1403_transpose_y_1 = const()[name = string("op_1403_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1403_cast_fp16 = matmul(transpose_x = var_1403_transpose_x_1, transpose_y = var_1403_transpose_y_1, x = q_13_cast_fp16, y = k_exp_27_cast_fp16)[name = string("op_1403_cast_fp16")];
            fp16 var_1404_to_fp16 = const()[name = string("op_1404_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_25_cast_fp16 = mul(x = var_1403_cast_fp16, y = var_1404_to_fp16)[name = string("attn_25_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_61_cast_fp16 = add(x = attn_25_cast_fp16, y = attention_mask_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_27_cast_fp16 = softmax(axis = var_1261, x = input_61_cast_fp16)[name = string("attn_27_cast_fp16")];
            bool out_13_transpose_x_0 = const()[name = string("out_13_transpose_x_0"), val = bool(false)];
            bool out_13_transpose_y_0 = const()[name = string("out_13_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_27_cast_fp16, y = v_exp_27_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<int32, [4]> var_1409_perm_0 = const()[name = string("op_1409_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1410 = const()[name = string("op_1410"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1409_cast_fp16 = transpose(perm = var_1409_perm_0, x = out_13_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 128, 2048]> input_63_cast_fp16 = reshape(shape = var_1410, x = var_1409_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_6_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98639616))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100736832))))[name = string("layers_6_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_o_proj_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_175_cast_fp16 = add(x = x_155_cast_fp16, y = linear_45_cast_fp16)[name = string("x_175_cast_fp16")];
            fp16 var_1260_promoted_3_to_fp16 = const()[name = string("op_1260_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1417_cast_fp16 = pow(x = x_175_cast_fp16, y = var_1260_promoted_3_to_fp16)[name = string("op_1417_cast_fp16")];
            tensor<int32, [1]> var_1419_axes_0 = const()[name = string("op_1419_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1419_keep_dims_0 = const()[name = string("op_1419_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1419_cast_fp16 = reduce_mean(axes = var_1419_axes_0, keep_dims = var_1419_keep_dims_0, x = var_1417_cast_fp16)[name = string("op_1419_cast_fp16")];
            fp16 var_1420_to_fp16 = const()[name = string("op_1420_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1421_cast_fp16 = add(x = var_1419_cast_fp16, y = var_1420_to_fp16)[name = string("op_1421_cast_fp16")];
            fp32 norm_55_epsilon_0 = const()[name = string("norm_55_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_55_cast_fp16 = rsqrt(epsilon = norm_55_epsilon_0, x = var_1421_cast_fp16)[name = string("norm_55_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1423_cast_fp16 = mul(x = x_175_cast_fp16, y = norm_55_cast_fp16)[name = string("op_1423_cast_fp16")];
            tensor<fp16, [1024]> layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100737408)))];
            tensor<fp16, [1, 128, 1024]> var_1424_cast_fp16 = mul(x = var_1423_cast_fp16, y = layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_1424_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_6_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100739520))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103885312))))[name = string("layers_6_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_gate_proj_weight_to_fp16_palettized, x = var_1424_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_1434_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_1434_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_6_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103885888))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107031680))))[name = string("layers_6_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_up_proj_weight_to_fp16_palettized, x = var_1424_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_69_cast_fp16 = mul(x = var_1434_cast_fp16, y = linear_47_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_6_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107032256))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110178048))))[name = string("layers_6_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_mlp_down_proj_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_181_cast_fp16 = add(x = x_175_cast_fp16, y = linear_48_cast_fp16)[name = string("x_181_cast_fp16")];
            int32 var_1455 = const()[name = string("op_1455"), val = int32(-1)];
            fp16 var_1454_promoted_to_fp16 = const()[name = string("op_1454_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1464_cast_fp16 = pow(x = x_181_cast_fp16, y = var_1454_promoted_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<int32, [1]> var_1466_axes_0 = const()[name = string("op_1466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1466_keep_dims_0 = const()[name = string("op_1466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1466_cast_fp16 = reduce_mean(axes = var_1466_axes_0, keep_dims = var_1466_keep_dims_0, x = var_1464_cast_fp16)[name = string("op_1466_cast_fp16")];
            fp16 var_1467_to_fp16 = const()[name = string("op_1467_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1468_cast_fp16 = add(x = var_1466_cast_fp16, y = var_1467_to_fp16)[name = string("op_1468_cast_fp16")];
            fp32 norm_57_epsilon_0 = const()[name = string("norm_57_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_57_cast_fp16 = rsqrt(epsilon = norm_57_epsilon_0, x = var_1468_cast_fp16)[name = string("norm_57_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1470_cast_fp16 = mul(x = x_181_cast_fp16, y = norm_57_cast_fp16)[name = string("op_1470_cast_fp16")];
            tensor<fp16, [1024]> layers_7_input_layernorm_weight_to_fp16 = const()[name = string("layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110178624)))];
            tensor<fp16, [1, 128, 1024]> var_1471_cast_fp16 = mul(x = var_1470_cast_fp16, y = layers_7_input_layernorm_weight_to_fp16)[name = string("op_1471_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_7_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110180736))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112277952))))[name = string("layers_7_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_q_proj_weight_to_fp16_palettized, x = var_1471_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1487 = const()[name = string("op_1487"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_1488_cast_fp16 = reshape(shape = var_1487, x = linear_49_cast_fp16)[name = string("op_1488_cast_fp16")];
            tensor<int32, [4]> x_187_perm_0 = const()[name = string("x_187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112278528))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113327168))))[name = string("layers_7_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_k_proj_weight_to_fp16_palettized, x = var_1471_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1492 = const()[name = string("op_1492"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1493_cast_fp16 = reshape(shape = var_1492, x = linear_50_cast_fp16)[name = string("op_1493_cast_fp16")];
            tensor<int32, [4]> x_191_perm_0 = const()[name = string("x_191_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113327744))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114376384))))[name = string("layers_7_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_v_proj_weight_to_fp16_palettized, x = var_1471_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1497 = const()[name = string("op_1497"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1498_cast_fp16 = reshape(shape = var_1497, x = linear_51_cast_fp16)[name = string("op_1498_cast_fp16")];
            tensor<int32, [4]> transpose_63_perm_0 = const()[name = string("transpose_63_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_1454_promoted_1_to_fp16 = const()[name = string("op_1454_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_187_cast_fp16 = transpose(perm = x_187_perm_0, x = var_1488_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 16, 128, 128]> var_1502_cast_fp16 = pow(x = x_187_cast_fp16, y = var_1454_promoted_1_to_fp16)[name = string("op_1502_cast_fp16")];
            tensor<int32, [1]> var_1504_axes_0 = const()[name = string("op_1504_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1504_keep_dims_0 = const()[name = string("op_1504_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_1504_cast_fp16 = reduce_mean(axes = var_1504_axes_0, keep_dims = var_1504_keep_dims_0, x = var_1502_cast_fp16)[name = string("op_1504_cast_fp16")];
            fp16 var_1505_to_fp16 = const()[name = string("op_1505_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_1506_cast_fp16 = add(x = var_1504_cast_fp16, y = var_1505_to_fp16)[name = string("op_1506_cast_fp16")];
            fp32 norm_59_epsilon_0 = const()[name = string("norm_59_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_59_cast_fp16 = rsqrt(epsilon = norm_59_epsilon_0, x = var_1506_cast_fp16)[name = string("norm_59_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1508_cast_fp16 = mul(x = x_187_cast_fp16, y = norm_59_cast_fp16)[name = string("op_1508_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114376960)))];
            tensor<fp16, [1, 16, 128, 128]> var_1509_cast_fp16 = mul(x = var_1508_cast_fp16, y = layers_7_self_attn_q_norm_weight_to_fp16)[name = string("op_1509_cast_fp16")];
            fp16 var_1454_promoted_2_to_fp16 = const()[name = string("op_1454_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_191_cast_fp16 = transpose(perm = x_191_perm_0, x = var_1493_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 128, 128]> var_1513_cast_fp16 = pow(x = x_191_cast_fp16, y = var_1454_promoted_2_to_fp16)[name = string("op_1513_cast_fp16")];
            tensor<int32, [1]> var_1515_axes_0 = const()[name = string("op_1515_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1515_keep_dims_0 = const()[name = string("op_1515_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_1515_cast_fp16 = reduce_mean(axes = var_1515_axes_0, keep_dims = var_1515_keep_dims_0, x = var_1513_cast_fp16)[name = string("op_1515_cast_fp16")];
            fp16 var_1516_to_fp16 = const()[name = string("op_1516_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_1517_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_to_fp16)[name = string("op_1517_cast_fp16")];
            fp32 norm_61_epsilon_0 = const()[name = string("norm_61_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_61_cast_fp16 = rsqrt(epsilon = norm_61_epsilon_0, x = var_1517_cast_fp16)[name = string("norm_61_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_1519_cast_fp16 = mul(x = x_191_cast_fp16, y = norm_61_cast_fp16)[name = string("op_1519_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114377280)))];
            tensor<fp16, [1, 8, 128, 128]> var_1520_cast_fp16 = mul(x = var_1519_cast_fp16, y = layers_7_self_attn_k_norm_weight_to_fp16)[name = string("op_1520_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_1509_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, 64])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_1509_cast_fp16)[name = string("x2_29_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1541_cast_fp16 = mul(x = x1_29_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1542_cast_fp16 = mul(x = x2_29_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1542_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1543_cast_fp16 = sub(x = var_1541_cast_fp16, y = var_1542_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1544_cast_fp16 = mul(x = x2_29_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1544_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1545_cast_fp16 = mul(x = x1_29_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1545_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1546_cast_fp16 = add(x = var_1544_cast_fp16, y = var_1545_cast_fp16)[name = string("op_1546_cast_fp16")];
            bool q_15_interleave_0 = const()[name = string("q_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_15_cast_fp16 = concat(axis = var_1455, interleave = q_15_interleave_0, values = (var_1543_cast_fp16, var_1546_cast_fp16))[name = string("q_15_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_1520_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, 64])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_1520_cast_fp16)[name = string("x2_31_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1568_cast_fp16 = mul(x = x1_31_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1568_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1569_cast_fp16 = mul(x = x2_31_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1569_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1570_cast_fp16 = sub(x = var_1568_cast_fp16, y = var_1569_cast_fp16)[name = string("op_1570_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1571_cast_fp16 = mul(x = x2_31_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1571_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1572_cast_fp16 = mul(x = x1_31_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1572_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1573_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = string("op_1573_cast_fp16")];
            bool var_1575_interleave_0 = const()[name = string("op_1575_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1575_cast_fp16 = concat(axis = var_1455, interleave = var_1575_interleave_0, values = (var_1570_cast_fp16, var_1573_cast_fp16))[name = string("op_1575_cast_fp16")];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_130 = const()[name = string("concat_130"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = var_1575_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [128, 1024]> reshape_43_cast_fp16 = reshape(shape = concat_130, x = transpose_29_cast_fp16)[name = string("reshape_43_cast_fp16")];
            bool matmul_14_transpose_x_1 = const()[name = string("matmul_14_transpose_x_1"), val = bool(true)];
            bool matmul_14_transpose_y_1 = const()[name = string("matmul_14_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_14_cast_fp16 = matmul(transpose_x = matmul_14_transpose_x_1, transpose_y = matmul_14_transpose_y_1, x = var_66_to_fp16, y = reshape_43_cast_fp16)[name = string("matmul_14_cast_fp16")];
            tensor<int32, [4]> concat_133 = const()[name = string("concat_133"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_44_cast_fp16 = reshape(shape = concat_133, x = matmul_14_cast_fp16)[name = string("reshape_44_cast_fp16")];
            tensor<int32, [4]> scattered_k_15_perm_0 = const()[name = string("scattered_k_15_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_138 = const()[name = string("concat_138"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_63_cast_fp16 = transpose(perm = transpose_63_perm_0, x = var_1498_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [128, 1024]> reshape_46_cast_fp16 = reshape(shape = concat_138, x = transpose_63_cast_fp16)[name = string("reshape_46_cast_fp16")];
            bool matmul_15_transpose_x_1 = const()[name = string("matmul_15_transpose_x_1"), val = bool(true)];
            bool matmul_15_transpose_y_1 = const()[name = string("matmul_15_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_15_cast_fp16 = matmul(transpose_x = matmul_15_transpose_x_1, transpose_y = matmul_15_transpose_y_1, x = var_66_to_fp16, y = reshape_46_cast_fp16)[name = string("matmul_15_cast_fp16")];
            tensor<int32, [4]> concat_141 = const()[name = string("concat_141"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_47_cast_fp16 = reshape(shape = concat_141, x = matmul_15_cast_fp16)[name = string("reshape_47_cast_fp16")];
            tensor<int32, [4]> scattered_v_15_perm_0 = const()[name = string("scattered_v_15_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_14 = read_state(input = k_cache_7)[name = string("read_state_14")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_45_cast_fp16 = mul(x = read_state_14, y = var_222_cast_fp16)[name = string("k_cache_45_cast_fp16")];
            write_state(data = k_cache_45_cast_fp16, input = k_cache_7)[name = string("coreml_update_state_84_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_84 = read_state(input = k_cache_7)[name = string("coreml_update_state_84")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_15_cast_fp16 = transpose(perm = scattered_k_15_perm_0, x = reshape_44_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_47_cast_fp16 = add(x = coreml_update_state_84, y = scattered_k_15_cast_fp16)[name = string("k_cache_47_cast_fp16")];
            write_state(data = k_cache_47_cast_fp16, input = k_cache_7)[name = string("coreml_update_state_85_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_85 = read_state(input = k_cache_7)[name = string("coreml_update_state_85")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_15 = read_state(input = v_cache_7)[name = string("read_state_15")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_45_cast_fp16 = mul(x = read_state_15, y = var_222_cast_fp16)[name = string("v_cache_45_cast_fp16")];
            write_state(data = v_cache_45_cast_fp16, input = v_cache_7)[name = string("coreml_update_state_86_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_86 = read_state(input = v_cache_7)[name = string("coreml_update_state_86")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_15_cast_fp16 = transpose(perm = scattered_v_15_perm_0, x = reshape_47_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_47_cast_fp16 = add(x = coreml_update_state_86, y = scattered_v_15_cast_fp16)[name = string("v_cache_47_cast_fp16")];
            write_state(data = v_cache_47_cast_fp16, input = v_cache_7)[name = string("coreml_update_state_87_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_87 = read_state(input = v_cache_7)[name = string("coreml_update_state_87")];
            tensor<int32, [1]> var_1586_axes_0 = const()[name = string("op_1586_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1586_cast_fp16 = expand_dims(axes = var_1586_axes_0, x = coreml_update_state_85)[name = string("op_1586_cast_fp16")];
            tensor<int32, [5]> k_exp_29_reps_0 = const()[name = string("k_exp_29_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_29_cast_fp16 = tile(reps = k_exp_29_reps_0, x = var_1586_cast_fp16)[name = string("k_exp_29_cast_fp16")];
            tensor<int32, [4]> var_1589 = const()[name = string("op_1589"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_31_cast_fp16 = reshape(shape = var_1589, x = k_exp_29_cast_fp16)[name = string("k_exp_31_cast_fp16")];
            tensor<int32, [1]> var_1591_axes_0 = const()[name = string("op_1591_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1591_cast_fp16 = expand_dims(axes = var_1591_axes_0, x = coreml_update_state_87)[name = string("op_1591_cast_fp16")];
            tensor<int32, [5]> v_exp_29_reps_0 = const()[name = string("v_exp_29_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_29_cast_fp16 = tile(reps = v_exp_29_reps_0, x = var_1591_cast_fp16)[name = string("v_exp_29_cast_fp16")];
            tensor<int32, [4]> var_1594 = const()[name = string("op_1594"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_31_cast_fp16 = reshape(shape = var_1594, x = v_exp_29_cast_fp16)[name = string("v_exp_31_cast_fp16")];
            bool var_1597_transpose_x_1 = const()[name = string("op_1597_transpose_x_1"), val = bool(false)];
            bool var_1597_transpose_y_1 = const()[name = string("op_1597_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1597_cast_fp16 = matmul(transpose_x = var_1597_transpose_x_1, transpose_y = var_1597_transpose_y_1, x = q_15_cast_fp16, y = k_exp_31_cast_fp16)[name = string("op_1597_cast_fp16")];
            fp16 var_1598_to_fp16 = const()[name = string("op_1598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_29_cast_fp16 = mul(x = var_1597_cast_fp16, y = var_1598_to_fp16)[name = string("attn_29_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_71_cast_fp16 = add(x = attn_29_cast_fp16, y = attention_mask_to_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_31_cast_fp16 = softmax(axis = var_1455, x = input_71_cast_fp16)[name = string("attn_31_cast_fp16")];
            bool out_15_transpose_x_0 = const()[name = string("out_15_transpose_x_0"), val = bool(false)];
            bool out_15_transpose_y_0 = const()[name = string("out_15_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_31_cast_fp16, y = v_exp_31_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<int32, [4]> var_1603_perm_0 = const()[name = string("op_1603_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1604 = const()[name = string("op_1604"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1603_cast_fp16 = transpose(perm = var_1603_perm_0, x = out_15_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 128, 2048]> input_73_cast_fp16 = reshape(shape = var_1604, x = var_1603_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_7_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114377600))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116474816))))[name = string("layers_7_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_201_cast_fp16 = add(x = x_181_cast_fp16, y = linear_52_cast_fp16)[name = string("x_201_cast_fp16")];
            fp16 var_1454_promoted_3_to_fp16 = const()[name = string("op_1454_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1611_cast_fp16 = pow(x = x_201_cast_fp16, y = var_1454_promoted_3_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> var_1613_axes_0 = const()[name = string("op_1613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1613_keep_dims_0 = const()[name = string("op_1613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1613_cast_fp16 = reduce_mean(axes = var_1613_axes_0, keep_dims = var_1613_keep_dims_0, x = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp16 var_1614_to_fp16 = const()[name = string("op_1614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1615_cast_fp16 = add(x = var_1613_cast_fp16, y = var_1614_to_fp16)[name = string("op_1615_cast_fp16")];
            fp32 norm_63_epsilon_0 = const()[name = string("norm_63_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_63_cast_fp16 = rsqrt(epsilon = norm_63_epsilon_0, x = var_1615_cast_fp16)[name = string("norm_63_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1617_cast_fp16 = mul(x = x_201_cast_fp16, y = norm_63_cast_fp16)[name = string("op_1617_cast_fp16")];
            tensor<fp16, [1024]> layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116475392)))];
            tensor<fp16, [1, 128, 1024]> var_1618_cast_fp16 = mul(x = var_1617_cast_fp16, y = layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_1618_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_7_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116477504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119623296))))[name = string("layers_7_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_gate_proj_weight_to_fp16_palettized, x = var_1618_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_1628_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_7_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119623872))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122769664))))[name = string("layers_7_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_up_proj_weight_to_fp16_palettized, x = var_1618_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_79_cast_fp16 = mul(x = var_1628_cast_fp16, y = linear_54_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_7_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122770240))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125916032))))[name = string("layers_7_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_mlp_down_proj_weight_to_fp16_palettized, x = input_79_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_207_cast_fp16 = add(x = x_201_cast_fp16, y = linear_55_cast_fp16)[name = string("x_207_cast_fp16")];
            int32 var_1649 = const()[name = string("op_1649"), val = int32(-1)];
            fp16 var_1648_promoted_to_fp16 = const()[name = string("op_1648_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1658_cast_fp16 = pow(x = x_207_cast_fp16, y = var_1648_promoted_to_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [1]> var_1660_axes_0 = const()[name = string("op_1660_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1660_keep_dims_0 = const()[name = string("op_1660_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1660_cast_fp16 = reduce_mean(axes = var_1660_axes_0, keep_dims = var_1660_keep_dims_0, x = var_1658_cast_fp16)[name = string("op_1660_cast_fp16")];
            fp16 var_1661_to_fp16 = const()[name = string("op_1661_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1662_cast_fp16 = add(x = var_1660_cast_fp16, y = var_1661_to_fp16)[name = string("op_1662_cast_fp16")];
            fp32 norm_65_epsilon_0 = const()[name = string("norm_65_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_65_cast_fp16 = rsqrt(epsilon = norm_65_epsilon_0, x = var_1662_cast_fp16)[name = string("norm_65_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1664_cast_fp16 = mul(x = x_207_cast_fp16, y = norm_65_cast_fp16)[name = string("op_1664_cast_fp16")];
            tensor<fp16, [1024]> layers_8_input_layernorm_weight_to_fp16 = const()[name = string("layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125916608)))];
            tensor<fp16, [1, 128, 1024]> var_1665_cast_fp16 = mul(x = var_1664_cast_fp16, y = layers_8_input_layernorm_weight_to_fp16)[name = string("op_1665_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_8_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125918720))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128015936))))[name = string("layers_8_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_q_proj_weight_to_fp16_palettized, x = var_1665_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1681 = const()[name = string("op_1681"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_1682_cast_fp16 = reshape(shape = var_1681, x = linear_56_cast_fp16)[name = string("op_1682_cast_fp16")];
            tensor<int32, [4]> x_213_perm_0 = const()[name = string("x_213_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128016512))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129065152))))[name = string("layers_8_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_k_proj_weight_to_fp16_palettized, x = var_1665_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1686 = const()[name = string("op_1686"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1687_cast_fp16 = reshape(shape = var_1686, x = linear_57_cast_fp16)[name = string("op_1687_cast_fp16")];
            tensor<int32, [4]> x_217_perm_0 = const()[name = string("x_217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129065728))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130114368))))[name = string("layers_8_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_v_proj_weight_to_fp16_palettized, x = var_1665_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1691 = const()[name = string("op_1691"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1692_cast_fp16 = reshape(shape = var_1691, x = linear_58_cast_fp16)[name = string("op_1692_cast_fp16")];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_1648_promoted_1_to_fp16 = const()[name = string("op_1648_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_213_cast_fp16 = transpose(perm = x_213_perm_0, x = var_1682_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 16, 128, 128]> var_1696_cast_fp16 = pow(x = x_213_cast_fp16, y = var_1648_promoted_1_to_fp16)[name = string("op_1696_cast_fp16")];
            tensor<int32, [1]> var_1698_axes_0 = const()[name = string("op_1698_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1698_keep_dims_0 = const()[name = string("op_1698_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_1698_cast_fp16 = reduce_mean(axes = var_1698_axes_0, keep_dims = var_1698_keep_dims_0, x = var_1696_cast_fp16)[name = string("op_1698_cast_fp16")];
            fp16 var_1699_to_fp16 = const()[name = string("op_1699_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_1700_cast_fp16 = add(x = var_1698_cast_fp16, y = var_1699_to_fp16)[name = string("op_1700_cast_fp16")];
            fp32 norm_67_epsilon_0 = const()[name = string("norm_67_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_67_cast_fp16 = rsqrt(epsilon = norm_67_epsilon_0, x = var_1700_cast_fp16)[name = string("norm_67_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1702_cast_fp16 = mul(x = x_213_cast_fp16, y = norm_67_cast_fp16)[name = string("op_1702_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130114944)))];
            tensor<fp16, [1, 16, 128, 128]> var_1703_cast_fp16 = mul(x = var_1702_cast_fp16, y = layers_8_self_attn_q_norm_weight_to_fp16)[name = string("op_1703_cast_fp16")];
            fp16 var_1648_promoted_2_to_fp16 = const()[name = string("op_1648_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_217_cast_fp16 = transpose(perm = x_217_perm_0, x = var_1687_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 8, 128, 128]> var_1707_cast_fp16 = pow(x = x_217_cast_fp16, y = var_1648_promoted_2_to_fp16)[name = string("op_1707_cast_fp16")];
            tensor<int32, [1]> var_1709_axes_0 = const()[name = string("op_1709_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1709_keep_dims_0 = const()[name = string("op_1709_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_1709_cast_fp16 = reduce_mean(axes = var_1709_axes_0, keep_dims = var_1709_keep_dims_0, x = var_1707_cast_fp16)[name = string("op_1709_cast_fp16")];
            fp16 var_1710_to_fp16 = const()[name = string("op_1710_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_1711_cast_fp16 = add(x = var_1709_cast_fp16, y = var_1710_to_fp16)[name = string("op_1711_cast_fp16")];
            fp32 norm_69_epsilon_0 = const()[name = string("norm_69_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_69_cast_fp16 = rsqrt(epsilon = norm_69_epsilon_0, x = var_1711_cast_fp16)[name = string("norm_69_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_1713_cast_fp16 = mul(x = x_217_cast_fp16, y = norm_69_cast_fp16)[name = string("op_1713_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130115264)))];
            tensor<fp16, [1, 8, 128, 128]> var_1714_cast_fp16 = mul(x = var_1713_cast_fp16, y = layers_8_self_attn_k_norm_weight_to_fp16)[name = string("op_1714_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_1703_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, 64])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_1703_cast_fp16)[name = string("x2_33_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1735_cast_fp16 = mul(x = x1_33_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1735_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1736_cast_fp16 = mul(x = x2_33_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1736_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1737_cast_fp16 = sub(x = var_1735_cast_fp16, y = var_1736_cast_fp16)[name = string("op_1737_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1738_cast_fp16 = mul(x = x2_33_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1739_cast_fp16 = mul(x = x1_33_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1739_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1740_cast_fp16 = add(x = var_1738_cast_fp16, y = var_1739_cast_fp16)[name = string("op_1740_cast_fp16")];
            bool q_17_interleave_0 = const()[name = string("q_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_17_cast_fp16 = concat(axis = var_1649, interleave = q_17_interleave_0, values = (var_1737_cast_fp16, var_1740_cast_fp16))[name = string("q_17_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_1714_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, 64])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_1714_cast_fp16)[name = string("x2_35_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1762_cast_fp16 = mul(x = x1_35_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1762_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1763_cast_fp16 = mul(x = x2_35_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1763_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1764_cast_fp16 = sub(x = var_1762_cast_fp16, y = var_1763_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1765_cast_fp16 = mul(x = x2_35_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1765_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1766_cast_fp16 = mul(x = x1_35_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1766_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1767_cast_fp16 = add(x = var_1765_cast_fp16, y = var_1766_cast_fp16)[name = string("op_1767_cast_fp16")];
            bool var_1769_interleave_0 = const()[name = string("op_1769_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1769_cast_fp16 = concat(axis = var_1649, interleave = var_1769_interleave_0, values = (var_1764_cast_fp16, var_1767_cast_fp16))[name = string("op_1769_cast_fp16")];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_148 = const()[name = string("concat_148"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_33_cast_fp16 = transpose(perm = transpose_33_perm_0, x = var_1769_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [128, 1024]> reshape_49_cast_fp16 = reshape(shape = concat_148, x = transpose_33_cast_fp16)[name = string("reshape_49_cast_fp16")];
            bool matmul_16_transpose_x_1 = const()[name = string("matmul_16_transpose_x_1"), val = bool(true)];
            bool matmul_16_transpose_y_1 = const()[name = string("matmul_16_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_16_cast_fp16 = matmul(transpose_x = matmul_16_transpose_x_1, transpose_y = matmul_16_transpose_y_1, x = var_66_to_fp16, y = reshape_49_cast_fp16)[name = string("matmul_16_cast_fp16")];
            tensor<int32, [4]> concat_151 = const()[name = string("concat_151"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_50_cast_fp16 = reshape(shape = concat_151, x = matmul_16_cast_fp16)[name = string("reshape_50_cast_fp16")];
            tensor<int32, [4]> scattered_k_17_perm_0 = const()[name = string("scattered_k_17_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_156 = const()[name = string("concat_156"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_64_cast_fp16 = transpose(perm = transpose_64_perm_0, x = var_1692_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [128, 1024]> reshape_52_cast_fp16 = reshape(shape = concat_156, x = transpose_64_cast_fp16)[name = string("reshape_52_cast_fp16")];
            bool matmul_17_transpose_x_1 = const()[name = string("matmul_17_transpose_x_1"), val = bool(true)];
            bool matmul_17_transpose_y_1 = const()[name = string("matmul_17_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_17_cast_fp16 = matmul(transpose_x = matmul_17_transpose_x_1, transpose_y = matmul_17_transpose_y_1, x = var_66_to_fp16, y = reshape_52_cast_fp16)[name = string("matmul_17_cast_fp16")];
            tensor<int32, [4]> concat_159 = const()[name = string("concat_159"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_53_cast_fp16 = reshape(shape = concat_159, x = matmul_17_cast_fp16)[name = string("reshape_53_cast_fp16")];
            tensor<int32, [4]> scattered_v_17_perm_0 = const()[name = string("scattered_v_17_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_16 = read_state(input = k_cache_8)[name = string("read_state_16")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_51_cast_fp16 = mul(x = read_state_16, y = var_222_cast_fp16)[name = string("k_cache_51_cast_fp16")];
            write_state(data = k_cache_51_cast_fp16, input = k_cache_8)[name = string("coreml_update_state_88_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_88 = read_state(input = k_cache_8)[name = string("coreml_update_state_88")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_17_cast_fp16 = transpose(perm = scattered_k_17_perm_0, x = reshape_50_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_53_cast_fp16 = add(x = coreml_update_state_88, y = scattered_k_17_cast_fp16)[name = string("k_cache_53_cast_fp16")];
            write_state(data = k_cache_53_cast_fp16, input = k_cache_8)[name = string("coreml_update_state_89_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_89 = read_state(input = k_cache_8)[name = string("coreml_update_state_89")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_17 = read_state(input = v_cache_8)[name = string("read_state_17")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_51_cast_fp16 = mul(x = read_state_17, y = var_222_cast_fp16)[name = string("v_cache_51_cast_fp16")];
            write_state(data = v_cache_51_cast_fp16, input = v_cache_8)[name = string("coreml_update_state_90_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_90 = read_state(input = v_cache_8)[name = string("coreml_update_state_90")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_17_cast_fp16 = transpose(perm = scattered_v_17_perm_0, x = reshape_53_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_53_cast_fp16 = add(x = coreml_update_state_90, y = scattered_v_17_cast_fp16)[name = string("v_cache_53_cast_fp16")];
            write_state(data = v_cache_53_cast_fp16, input = v_cache_8)[name = string("coreml_update_state_91_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_91 = read_state(input = v_cache_8)[name = string("coreml_update_state_91")];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = string("op_1780_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1780_cast_fp16 = expand_dims(axes = var_1780_axes_0, x = coreml_update_state_89)[name = string("op_1780_cast_fp16")];
            tensor<int32, [5]> k_exp_33_reps_0 = const()[name = string("k_exp_33_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_33_cast_fp16 = tile(reps = k_exp_33_reps_0, x = var_1780_cast_fp16)[name = string("k_exp_33_cast_fp16")];
            tensor<int32, [4]> var_1783 = const()[name = string("op_1783"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_35_cast_fp16 = reshape(shape = var_1783, x = k_exp_33_cast_fp16)[name = string("k_exp_35_cast_fp16")];
            tensor<int32, [1]> var_1785_axes_0 = const()[name = string("op_1785_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1785_cast_fp16 = expand_dims(axes = var_1785_axes_0, x = coreml_update_state_91)[name = string("op_1785_cast_fp16")];
            tensor<int32, [5]> v_exp_33_reps_0 = const()[name = string("v_exp_33_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_33_cast_fp16 = tile(reps = v_exp_33_reps_0, x = var_1785_cast_fp16)[name = string("v_exp_33_cast_fp16")];
            tensor<int32, [4]> var_1788 = const()[name = string("op_1788"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_35_cast_fp16 = reshape(shape = var_1788, x = v_exp_33_cast_fp16)[name = string("v_exp_35_cast_fp16")];
            bool var_1791_transpose_x_1 = const()[name = string("op_1791_transpose_x_1"), val = bool(false)];
            bool var_1791_transpose_y_1 = const()[name = string("op_1791_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1791_cast_fp16 = matmul(transpose_x = var_1791_transpose_x_1, transpose_y = var_1791_transpose_y_1, x = q_17_cast_fp16, y = k_exp_35_cast_fp16)[name = string("op_1791_cast_fp16")];
            fp16 var_1792_to_fp16 = const()[name = string("op_1792_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_33_cast_fp16 = mul(x = var_1791_cast_fp16, y = var_1792_to_fp16)[name = string("attn_33_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_81_cast_fp16 = add(x = attn_33_cast_fp16, y = attention_mask_to_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_35_cast_fp16 = softmax(axis = var_1649, x = input_81_cast_fp16)[name = string("attn_35_cast_fp16")];
            bool out_17_transpose_x_0 = const()[name = string("out_17_transpose_x_0"), val = bool(false)];
            bool out_17_transpose_y_0 = const()[name = string("out_17_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_35_cast_fp16, y = v_exp_35_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<int32, [4]> var_1797_perm_0 = const()[name = string("op_1797_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1798 = const()[name = string("op_1798"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1797_cast_fp16 = transpose(perm = var_1797_perm_0, x = out_17_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 128, 2048]> input_83_cast_fp16 = reshape(shape = var_1798, x = var_1797_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_8_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130115584))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132212800))))[name = string("layers_8_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_o_proj_weight_to_fp16_palettized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_227_cast_fp16 = add(x = x_207_cast_fp16, y = linear_59_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_1648_promoted_3_to_fp16 = const()[name = string("op_1648_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1805_cast_fp16 = pow(x = x_227_cast_fp16, y = var_1648_promoted_3_to_fp16)[name = string("op_1805_cast_fp16")];
            tensor<int32, [1]> var_1807_axes_0 = const()[name = string("op_1807_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1807_keep_dims_0 = const()[name = string("op_1807_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1807_cast_fp16 = reduce_mean(axes = var_1807_axes_0, keep_dims = var_1807_keep_dims_0, x = var_1805_cast_fp16)[name = string("op_1807_cast_fp16")];
            fp16 var_1808_to_fp16 = const()[name = string("op_1808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1809_cast_fp16 = add(x = var_1807_cast_fp16, y = var_1808_to_fp16)[name = string("op_1809_cast_fp16")];
            fp32 norm_71_epsilon_0 = const()[name = string("norm_71_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_71_cast_fp16 = rsqrt(epsilon = norm_71_epsilon_0, x = var_1809_cast_fp16)[name = string("norm_71_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1811_cast_fp16 = mul(x = x_227_cast_fp16, y = norm_71_cast_fp16)[name = string("op_1811_cast_fp16")];
            tensor<fp16, [1024]> layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132213376)))];
            tensor<fp16, [1, 128, 1024]> var_1812_cast_fp16 = mul(x = var_1811_cast_fp16, y = layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_1812_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_8_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132215488))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135361280))))[name = string("layers_8_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_gate_proj_weight_to_fp16_palettized, x = var_1812_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_1822_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_1822_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_8_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135361856))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138507648))))[name = string("layers_8_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_up_proj_weight_to_fp16_palettized, x = var_1812_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_89_cast_fp16 = mul(x = var_1822_cast_fp16, y = linear_61_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_8_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138508224))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141654016))))[name = string("layers_8_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_mlp_down_proj_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_233_cast_fp16 = add(x = x_227_cast_fp16, y = linear_62_cast_fp16)[name = string("x_233_cast_fp16")];
            int32 var_1843 = const()[name = string("op_1843"), val = int32(-1)];
            fp16 var_1842_promoted_to_fp16 = const()[name = string("op_1842_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1852_cast_fp16 = pow(x = x_233_cast_fp16, y = var_1842_promoted_to_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [1]> var_1854_axes_0 = const()[name = string("op_1854_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1854_keep_dims_0 = const()[name = string("op_1854_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_1854_cast_fp16 = reduce_mean(axes = var_1854_axes_0, keep_dims = var_1854_keep_dims_0, x = var_1852_cast_fp16)[name = string("op_1854_cast_fp16")];
            fp16 var_1855_to_fp16 = const()[name = string("op_1855_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_1856_cast_fp16 = add(x = var_1854_cast_fp16, y = var_1855_to_fp16)[name = string("op_1856_cast_fp16")];
            fp32 norm_73_epsilon_0 = const()[name = string("norm_73_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_73_cast_fp16 = rsqrt(epsilon = norm_73_epsilon_0, x = var_1856_cast_fp16)[name = string("norm_73_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_1858_cast_fp16 = mul(x = x_233_cast_fp16, y = norm_73_cast_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1024]> layers_9_input_layernorm_weight_to_fp16 = const()[name = string("layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141654592)))];
            tensor<fp16, [1, 128, 1024]> var_1859_cast_fp16 = mul(x = var_1858_cast_fp16, y = layers_9_input_layernorm_weight_to_fp16)[name = string("op_1859_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_9_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141656704))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143753920))))[name = string("layers_9_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_q_proj_weight_to_fp16_palettized, x = var_1859_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1875 = const()[name = string("op_1875"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_1876_cast_fp16 = reshape(shape = var_1875, x = linear_63_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<int32, [4]> x_239_perm_0 = const()[name = string("x_239_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143754496))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144803136))))[name = string("layers_9_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_k_proj_weight_to_fp16_palettized, x = var_1859_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1880 = const()[name = string("op_1880"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1881_cast_fp16 = reshape(shape = var_1880, x = linear_64_cast_fp16)[name = string("op_1881_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144803712))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145852352))))[name = string("layers_9_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_v_proj_weight_to_fp16_palettized, x = var_1859_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1885 = const()[name = string("op_1885"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_1886_cast_fp16 = reshape(shape = var_1885, x = linear_65_cast_fp16)[name = string("op_1886_cast_fp16")];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_1842_promoted_1_to_fp16 = const()[name = string("op_1842_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_239_cast_fp16 = transpose(perm = x_239_perm_0, x = var_1876_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 128, 128]> var_1890_cast_fp16 = pow(x = x_239_cast_fp16, y = var_1842_promoted_1_to_fp16)[name = string("op_1890_cast_fp16")];
            tensor<int32, [1]> var_1892_axes_0 = const()[name = string("op_1892_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1892_keep_dims_0 = const()[name = string("op_1892_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_1892_cast_fp16 = reduce_mean(axes = var_1892_axes_0, keep_dims = var_1892_keep_dims_0, x = var_1890_cast_fp16)[name = string("op_1892_cast_fp16")];
            fp16 var_1893_to_fp16 = const()[name = string("op_1893_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_1894_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_to_fp16)[name = string("op_1894_cast_fp16")];
            fp32 norm_75_epsilon_0 = const()[name = string("norm_75_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_75_cast_fp16 = rsqrt(epsilon = norm_75_epsilon_0, x = var_1894_cast_fp16)[name = string("norm_75_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_1896_cast_fp16 = mul(x = x_239_cast_fp16, y = norm_75_cast_fp16)[name = string("op_1896_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145852928)))];
            tensor<fp16, [1, 16, 128, 128]> var_1897_cast_fp16 = mul(x = var_1896_cast_fp16, y = layers_9_self_attn_q_norm_weight_to_fp16)[name = string("op_1897_cast_fp16")];
            fp16 var_1842_promoted_2_to_fp16 = const()[name = string("op_1842_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_1881_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 128, 128]> var_1901_cast_fp16 = pow(x = x_243_cast_fp16, y = var_1842_promoted_2_to_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1903_axes_0 = const()[name = string("op_1903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1903_keep_dims_0 = const()[name = string("op_1903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_1903_cast_fp16 = reduce_mean(axes = var_1903_axes_0, keep_dims = var_1903_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1903_cast_fp16")];
            fp16 var_1904_to_fp16 = const()[name = string("op_1904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_1905_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1904_to_fp16)[name = string("op_1905_cast_fp16")];
            fp32 norm_77_epsilon_0 = const()[name = string("norm_77_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_77_cast_fp16 = rsqrt(epsilon = norm_77_epsilon_0, x = var_1905_cast_fp16)[name = string("norm_77_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_1907_cast_fp16 = mul(x = x_243_cast_fp16, y = norm_77_cast_fp16)[name = string("op_1907_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145853248)))];
            tensor<fp16, [1, 8, 128, 128]> var_1908_cast_fp16 = mul(x = var_1907_cast_fp16, y = layers_9_self_attn_k_norm_weight_to_fp16)[name = string("op_1908_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_1897_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, 64])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_1897_cast_fp16)[name = string("x2_37_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1929_cast_fp16 = mul(x = x1_37_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1929_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1930_cast_fp16 = mul(x = x2_37_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1930_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1931_cast_fp16 = sub(x = var_1929_cast_fp16, y = var_1930_cast_fp16)[name = string("op_1931_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1932_cast_fp16 = mul(x = x2_37_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1932_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1933_cast_fp16 = mul(x = x1_37_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_1934_cast_fp16 = add(x = var_1932_cast_fp16, y = var_1933_cast_fp16)[name = string("op_1934_cast_fp16")];
            bool q_19_interleave_0 = const()[name = string("q_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_19_cast_fp16 = concat(axis = var_1843, interleave = q_19_interleave_0, values = (var_1931_cast_fp16, var_1934_cast_fp16))[name = string("q_19_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_1908_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, 64])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_1908_cast_fp16)[name = string("x2_39_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1956_cast_fp16 = mul(x = x1_39_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1956_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1957_cast_fp16 = mul(x = x2_39_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1957_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1958_cast_fp16 = sub(x = var_1956_cast_fp16, y = var_1957_cast_fp16)[name = string("op_1958_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1959_cast_fp16 = mul(x = x2_39_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_1959_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1960_cast_fp16 = mul(x = x1_39_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_1960_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_1961_cast_fp16 = add(x = var_1959_cast_fp16, y = var_1960_cast_fp16)[name = string("op_1961_cast_fp16")];
            bool var_1963_interleave_0 = const()[name = string("op_1963_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_1963_cast_fp16 = concat(axis = var_1843, interleave = var_1963_interleave_0, values = (var_1958_cast_fp16, var_1961_cast_fp16))[name = string("op_1963_cast_fp16")];
            tensor<int32, [4]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_166 = const()[name = string("concat_166"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_37_cast_fp16 = transpose(perm = transpose_37_perm_0, x = var_1963_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [128, 1024]> reshape_55_cast_fp16 = reshape(shape = concat_166, x = transpose_37_cast_fp16)[name = string("reshape_55_cast_fp16")];
            bool matmul_18_transpose_x_1 = const()[name = string("matmul_18_transpose_x_1"), val = bool(true)];
            bool matmul_18_transpose_y_1 = const()[name = string("matmul_18_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_18_cast_fp16 = matmul(transpose_x = matmul_18_transpose_x_1, transpose_y = matmul_18_transpose_y_1, x = var_66_to_fp16, y = reshape_55_cast_fp16)[name = string("matmul_18_cast_fp16")];
            tensor<int32, [4]> concat_169 = const()[name = string("concat_169"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_56_cast_fp16 = reshape(shape = concat_169, x = matmul_18_cast_fp16)[name = string("reshape_56_cast_fp16")];
            tensor<int32, [4]> scattered_k_19_perm_0 = const()[name = string("scattered_k_19_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_174 = const()[name = string("concat_174"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_65_cast_fp16 = transpose(perm = transpose_65_perm_0, x = var_1886_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [128, 1024]> reshape_58_cast_fp16 = reshape(shape = concat_174, x = transpose_65_cast_fp16)[name = string("reshape_58_cast_fp16")];
            bool matmul_19_transpose_x_1 = const()[name = string("matmul_19_transpose_x_1"), val = bool(true)];
            bool matmul_19_transpose_y_1 = const()[name = string("matmul_19_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_19_cast_fp16 = matmul(transpose_x = matmul_19_transpose_x_1, transpose_y = matmul_19_transpose_y_1, x = var_66_to_fp16, y = reshape_58_cast_fp16)[name = string("matmul_19_cast_fp16")];
            tensor<int32, [4]> concat_177 = const()[name = string("concat_177"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_59_cast_fp16 = reshape(shape = concat_177, x = matmul_19_cast_fp16)[name = string("reshape_59_cast_fp16")];
            tensor<int32, [4]> scattered_v_19_perm_0 = const()[name = string("scattered_v_19_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_18 = read_state(input = k_cache_9)[name = string("read_state_18")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_57_cast_fp16 = mul(x = read_state_18, y = var_222_cast_fp16)[name = string("k_cache_57_cast_fp16")];
            write_state(data = k_cache_57_cast_fp16, input = k_cache_9)[name = string("coreml_update_state_92_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_92 = read_state(input = k_cache_9)[name = string("coreml_update_state_92")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_19_cast_fp16 = transpose(perm = scattered_k_19_perm_0, x = reshape_56_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_59_cast_fp16 = add(x = coreml_update_state_92, y = scattered_k_19_cast_fp16)[name = string("k_cache_59_cast_fp16")];
            write_state(data = k_cache_59_cast_fp16, input = k_cache_9)[name = string("coreml_update_state_93_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_93 = read_state(input = k_cache_9)[name = string("coreml_update_state_93")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_19 = read_state(input = v_cache_9)[name = string("read_state_19")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_57_cast_fp16 = mul(x = read_state_19, y = var_222_cast_fp16)[name = string("v_cache_57_cast_fp16")];
            write_state(data = v_cache_57_cast_fp16, input = v_cache_9)[name = string("coreml_update_state_94_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_94 = read_state(input = v_cache_9)[name = string("coreml_update_state_94")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_19_cast_fp16 = transpose(perm = scattered_v_19_perm_0, x = reshape_59_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_59_cast_fp16 = add(x = coreml_update_state_94, y = scattered_v_19_cast_fp16)[name = string("v_cache_59_cast_fp16")];
            write_state(data = v_cache_59_cast_fp16, input = v_cache_9)[name = string("coreml_update_state_95_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_95 = read_state(input = v_cache_9)[name = string("coreml_update_state_95")];
            tensor<int32, [1]> var_1974_axes_0 = const()[name = string("op_1974_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1974_cast_fp16 = expand_dims(axes = var_1974_axes_0, x = coreml_update_state_93)[name = string("op_1974_cast_fp16")];
            tensor<int32, [5]> k_exp_37_reps_0 = const()[name = string("k_exp_37_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_37_cast_fp16 = tile(reps = k_exp_37_reps_0, x = var_1974_cast_fp16)[name = string("k_exp_37_cast_fp16")];
            tensor<int32, [4]> var_1977 = const()[name = string("op_1977"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_39_cast_fp16 = reshape(shape = var_1977, x = k_exp_37_cast_fp16)[name = string("k_exp_39_cast_fp16")];
            tensor<int32, [1]> var_1979_axes_0 = const()[name = string("op_1979_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_1979_cast_fp16 = expand_dims(axes = var_1979_axes_0, x = coreml_update_state_95)[name = string("op_1979_cast_fp16")];
            tensor<int32, [5]> v_exp_37_reps_0 = const()[name = string("v_exp_37_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_37_cast_fp16 = tile(reps = v_exp_37_reps_0, x = var_1979_cast_fp16)[name = string("v_exp_37_cast_fp16")];
            tensor<int32, [4]> var_1982 = const()[name = string("op_1982"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_39_cast_fp16 = reshape(shape = var_1982, x = v_exp_37_cast_fp16)[name = string("v_exp_39_cast_fp16")];
            bool var_1985_transpose_x_1 = const()[name = string("op_1985_transpose_x_1"), val = bool(false)];
            bool var_1985_transpose_y_1 = const()[name = string("op_1985_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_1985_cast_fp16 = matmul(transpose_x = var_1985_transpose_x_1, transpose_y = var_1985_transpose_y_1, x = q_19_cast_fp16, y = k_exp_39_cast_fp16)[name = string("op_1985_cast_fp16")];
            fp16 var_1986_to_fp16 = const()[name = string("op_1986_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_37_cast_fp16 = mul(x = var_1985_cast_fp16, y = var_1986_to_fp16)[name = string("attn_37_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_91_cast_fp16 = add(x = attn_37_cast_fp16, y = attention_mask_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_39_cast_fp16 = softmax(axis = var_1843, x = input_91_cast_fp16)[name = string("attn_39_cast_fp16")];
            bool out_19_transpose_x_0 = const()[name = string("out_19_transpose_x_0"), val = bool(false)];
            bool out_19_transpose_y_0 = const()[name = string("out_19_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_39_cast_fp16, y = v_exp_39_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<int32, [4]> var_1991_perm_0 = const()[name = string("op_1991_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1992 = const()[name = string("op_1992"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_1991_cast_fp16 = transpose(perm = var_1991_perm_0, x = out_19_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 128, 2048]> input_93_cast_fp16 = reshape(shape = var_1992, x = var_1991_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_9_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145853568))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147950784))))[name = string("layers_9_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_o_proj_weight_to_fp16_palettized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_253_cast_fp16 = add(x = x_233_cast_fp16, y = linear_66_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_1842_promoted_3_to_fp16 = const()[name = string("op_1842_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_1999_cast_fp16 = pow(x = x_253_cast_fp16, y = var_1842_promoted_3_to_fp16)[name = string("op_1999_cast_fp16")];
            tensor<int32, [1]> var_2001_axes_0 = const()[name = string("op_2001_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2001_keep_dims_0 = const()[name = string("op_2001_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2001_cast_fp16 = reduce_mean(axes = var_2001_axes_0, keep_dims = var_2001_keep_dims_0, x = var_1999_cast_fp16)[name = string("op_2001_cast_fp16")];
            fp16 var_2002_to_fp16 = const()[name = string("op_2002_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2003_cast_fp16 = add(x = var_2001_cast_fp16, y = var_2002_to_fp16)[name = string("op_2003_cast_fp16")];
            fp32 norm_79_epsilon_0 = const()[name = string("norm_79_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_79_cast_fp16 = rsqrt(epsilon = norm_79_epsilon_0, x = var_2003_cast_fp16)[name = string("norm_79_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2005_cast_fp16 = mul(x = x_253_cast_fp16, y = norm_79_cast_fp16)[name = string("op_2005_cast_fp16")];
            tensor<fp16, [1024]> layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147951360)))];
            tensor<fp16, [1, 128, 1024]> var_2006_cast_fp16 = mul(x = var_2005_cast_fp16, y = layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_2006_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_9_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147953472))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151099264))))[name = string("layers_9_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_gate_proj_weight_to_fp16_palettized, x = var_2006_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_2016_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_2016_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_9_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151099840))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154245632))))[name = string("layers_9_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_up_proj_weight_to_fp16_palettized, x = var_2006_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_99_cast_fp16 = mul(x = var_2016_cast_fp16, y = linear_68_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_9_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154246208))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157392000))))[name = string("layers_9_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_mlp_down_proj_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_259_cast_fp16 = add(x = x_253_cast_fp16, y = linear_69_cast_fp16)[name = string("x_259_cast_fp16")];
            int32 var_2037 = const()[name = string("op_2037"), val = int32(-1)];
            fp16 var_2036_promoted_to_fp16 = const()[name = string("op_2036_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2046_cast_fp16 = pow(x = x_259_cast_fp16, y = var_2036_promoted_to_fp16)[name = string("op_2046_cast_fp16")];
            tensor<int32, [1]> var_2048_axes_0 = const()[name = string("op_2048_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2048_keep_dims_0 = const()[name = string("op_2048_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2048_cast_fp16 = reduce_mean(axes = var_2048_axes_0, keep_dims = var_2048_keep_dims_0, x = var_2046_cast_fp16)[name = string("op_2048_cast_fp16")];
            fp16 var_2049_to_fp16 = const()[name = string("op_2049_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2050_cast_fp16 = add(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = string("op_2050_cast_fp16")];
            fp32 norm_81_epsilon_0 = const()[name = string("norm_81_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_81_cast_fp16 = rsqrt(epsilon = norm_81_epsilon_0, x = var_2050_cast_fp16)[name = string("norm_81_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2052_cast_fp16 = mul(x = x_259_cast_fp16, y = norm_81_cast_fp16)[name = string("op_2052_cast_fp16")];
            tensor<fp16, [1024]> layers_10_input_layernorm_weight_to_fp16 = const()[name = string("layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157392576)))];
            tensor<fp16, [1, 128, 1024]> var_2053_cast_fp16 = mul(x = var_2052_cast_fp16, y = layers_10_input_layernorm_weight_to_fp16)[name = string("op_2053_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_10_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157394688))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159491904))))[name = string("layers_10_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_q_proj_weight_to_fp16_palettized, x = var_2053_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_2069 = const()[name = string("op_2069"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_2070_cast_fp16 = reshape(shape = var_2069, x = linear_70_cast_fp16)[name = string("op_2070_cast_fp16")];
            tensor<int32, [4]> x_265_perm_0 = const()[name = string("x_265_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159492480))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160541120))))[name = string("layers_10_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_k_proj_weight_to_fp16_palettized, x = var_2053_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_2074 = const()[name = string("op_2074"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2075_cast_fp16 = reshape(shape = var_2074, x = linear_71_cast_fp16)[name = string("op_2075_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160541696))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161590336))))[name = string("layers_10_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_v_proj_weight_to_fp16_palettized, x = var_2053_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_2079 = const()[name = string("op_2079"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2080_cast_fp16 = reshape(shape = var_2079, x = linear_72_cast_fp16)[name = string("op_2080_cast_fp16")];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_2036_promoted_1_to_fp16 = const()[name = string("op_2036_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_265_cast_fp16 = transpose(perm = x_265_perm_0, x = var_2070_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 16, 128, 128]> var_2084_cast_fp16 = pow(x = x_265_cast_fp16, y = var_2036_promoted_1_to_fp16)[name = string("op_2084_cast_fp16")];
            tensor<int32, [1]> var_2086_axes_0 = const()[name = string("op_2086_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2086_keep_dims_0 = const()[name = string("op_2086_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_2086_cast_fp16 = reduce_mean(axes = var_2086_axes_0, keep_dims = var_2086_keep_dims_0, x = var_2084_cast_fp16)[name = string("op_2086_cast_fp16")];
            fp16 var_2087_to_fp16 = const()[name = string("op_2087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_2088_cast_fp16 = add(x = var_2086_cast_fp16, y = var_2087_to_fp16)[name = string("op_2088_cast_fp16")];
            fp32 norm_83_epsilon_0 = const()[name = string("norm_83_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_83_cast_fp16 = rsqrt(epsilon = norm_83_epsilon_0, x = var_2088_cast_fp16)[name = string("norm_83_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2090_cast_fp16 = mul(x = x_265_cast_fp16, y = norm_83_cast_fp16)[name = string("op_2090_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161590912)))];
            tensor<fp16, [1, 16, 128, 128]> var_2091_cast_fp16 = mul(x = var_2090_cast_fp16, y = layers_10_self_attn_q_norm_weight_to_fp16)[name = string("op_2091_cast_fp16")];
            fp16 var_2036_promoted_2_to_fp16 = const()[name = string("op_2036_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_2075_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 128, 128]> var_2095_cast_fp16 = pow(x = x_269_cast_fp16, y = var_2036_promoted_2_to_fp16)[name = string("op_2095_cast_fp16")];
            tensor<int32, [1]> var_2097_axes_0 = const()[name = string("op_2097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2097_keep_dims_0 = const()[name = string("op_2097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_2097_cast_fp16 = reduce_mean(axes = var_2097_axes_0, keep_dims = var_2097_keep_dims_0, x = var_2095_cast_fp16)[name = string("op_2097_cast_fp16")];
            fp16 var_2098_to_fp16 = const()[name = string("op_2098_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_2099_cast_fp16 = add(x = var_2097_cast_fp16, y = var_2098_to_fp16)[name = string("op_2099_cast_fp16")];
            fp32 norm_85_epsilon_0 = const()[name = string("norm_85_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_85_cast_fp16 = rsqrt(epsilon = norm_85_epsilon_0, x = var_2099_cast_fp16)[name = string("norm_85_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_2101_cast_fp16 = mul(x = x_269_cast_fp16, y = norm_85_cast_fp16)[name = string("op_2101_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161591232)))];
            tensor<fp16, [1, 8, 128, 128]> var_2102_cast_fp16 = mul(x = var_2101_cast_fp16, y = layers_10_self_attn_k_norm_weight_to_fp16)[name = string("op_2102_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_2091_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, 64])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_2091_cast_fp16)[name = string("x2_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2123_cast_fp16 = mul(x = x1_41_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2123_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2124_cast_fp16 = mul(x = x2_41_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2125_cast_fp16 = sub(x = var_2123_cast_fp16, y = var_2124_cast_fp16)[name = string("op_2125_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2126_cast_fp16 = mul(x = x2_41_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2126_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2127_cast_fp16 = mul(x = x1_41_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2127_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2128_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = string("op_2128_cast_fp16")];
            bool q_21_interleave_0 = const()[name = string("q_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_21_cast_fp16 = concat(axis = var_2037, interleave = q_21_interleave_0, values = (var_2125_cast_fp16, var_2128_cast_fp16))[name = string("q_21_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_2102_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, 64])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_2102_cast_fp16)[name = string("x2_43_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2150_cast_fp16 = mul(x = x1_43_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2150_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2151_cast_fp16 = mul(x = x2_43_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2151_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2152_cast_fp16 = sub(x = var_2150_cast_fp16, y = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2153_cast_fp16 = mul(x = x2_43_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2154_cast_fp16 = mul(x = x1_43_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2155_cast_fp16 = add(x = var_2153_cast_fp16, y = var_2154_cast_fp16)[name = string("op_2155_cast_fp16")];
            bool var_2157_interleave_0 = const()[name = string("op_2157_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2157_cast_fp16 = concat(axis = var_2037, interleave = var_2157_interleave_0, values = (var_2152_cast_fp16, var_2155_cast_fp16))[name = string("op_2157_cast_fp16")];
            tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_184 = const()[name = string("concat_184"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_41_cast_fp16 = transpose(perm = transpose_41_perm_0, x = var_2157_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [128, 1024]> reshape_61_cast_fp16 = reshape(shape = concat_184, x = transpose_41_cast_fp16)[name = string("reshape_61_cast_fp16")];
            bool matmul_20_transpose_x_1 = const()[name = string("matmul_20_transpose_x_1"), val = bool(true)];
            bool matmul_20_transpose_y_1 = const()[name = string("matmul_20_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_20_cast_fp16 = matmul(transpose_x = matmul_20_transpose_x_1, transpose_y = matmul_20_transpose_y_1, x = var_66_to_fp16, y = reshape_61_cast_fp16)[name = string("matmul_20_cast_fp16")];
            tensor<int32, [4]> concat_187 = const()[name = string("concat_187"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_62_cast_fp16 = reshape(shape = concat_187, x = matmul_20_cast_fp16)[name = string("reshape_62_cast_fp16")];
            tensor<int32, [4]> scattered_k_21_perm_0 = const()[name = string("scattered_k_21_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_192 = const()[name = string("concat_192"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_66_cast_fp16 = transpose(perm = transpose_66_perm_0, x = var_2080_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [128, 1024]> reshape_64_cast_fp16 = reshape(shape = concat_192, x = transpose_66_cast_fp16)[name = string("reshape_64_cast_fp16")];
            bool matmul_21_transpose_x_1 = const()[name = string("matmul_21_transpose_x_1"), val = bool(true)];
            bool matmul_21_transpose_y_1 = const()[name = string("matmul_21_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_21_cast_fp16 = matmul(transpose_x = matmul_21_transpose_x_1, transpose_y = matmul_21_transpose_y_1, x = var_66_to_fp16, y = reshape_64_cast_fp16)[name = string("matmul_21_cast_fp16")];
            tensor<int32, [4]> concat_195 = const()[name = string("concat_195"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_65_cast_fp16 = reshape(shape = concat_195, x = matmul_21_cast_fp16)[name = string("reshape_65_cast_fp16")];
            tensor<int32, [4]> scattered_v_21_perm_0 = const()[name = string("scattered_v_21_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_20 = read_state(input = k_cache_10)[name = string("read_state_20")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_63_cast_fp16 = mul(x = read_state_20, y = var_222_cast_fp16)[name = string("k_cache_63_cast_fp16")];
            write_state(data = k_cache_63_cast_fp16, input = k_cache_10)[name = string("coreml_update_state_96_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_96 = read_state(input = k_cache_10)[name = string("coreml_update_state_96")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_21_cast_fp16 = transpose(perm = scattered_k_21_perm_0, x = reshape_62_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_65_cast_fp16 = add(x = coreml_update_state_96, y = scattered_k_21_cast_fp16)[name = string("k_cache_65_cast_fp16")];
            write_state(data = k_cache_65_cast_fp16, input = k_cache_10)[name = string("coreml_update_state_97_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_97 = read_state(input = k_cache_10)[name = string("coreml_update_state_97")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_21 = read_state(input = v_cache_10)[name = string("read_state_21")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_63_cast_fp16 = mul(x = read_state_21, y = var_222_cast_fp16)[name = string("v_cache_63_cast_fp16")];
            write_state(data = v_cache_63_cast_fp16, input = v_cache_10)[name = string("coreml_update_state_98_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_98 = read_state(input = v_cache_10)[name = string("coreml_update_state_98")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_21_cast_fp16 = transpose(perm = scattered_v_21_perm_0, x = reshape_65_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_65_cast_fp16 = add(x = coreml_update_state_98, y = scattered_v_21_cast_fp16)[name = string("v_cache_65_cast_fp16")];
            write_state(data = v_cache_65_cast_fp16, input = v_cache_10)[name = string("coreml_update_state_99_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_99 = read_state(input = v_cache_10)[name = string("coreml_update_state_99")];
            tensor<int32, [1]> var_2168_axes_0 = const()[name = string("op_2168_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2168_cast_fp16 = expand_dims(axes = var_2168_axes_0, x = coreml_update_state_97)[name = string("op_2168_cast_fp16")];
            tensor<int32, [5]> k_exp_41_reps_0 = const()[name = string("k_exp_41_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_41_cast_fp16 = tile(reps = k_exp_41_reps_0, x = var_2168_cast_fp16)[name = string("k_exp_41_cast_fp16")];
            tensor<int32, [4]> var_2171 = const()[name = string("op_2171"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_43_cast_fp16 = reshape(shape = var_2171, x = k_exp_41_cast_fp16)[name = string("k_exp_43_cast_fp16")];
            tensor<int32, [1]> var_2173_axes_0 = const()[name = string("op_2173_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2173_cast_fp16 = expand_dims(axes = var_2173_axes_0, x = coreml_update_state_99)[name = string("op_2173_cast_fp16")];
            tensor<int32, [5]> v_exp_41_reps_0 = const()[name = string("v_exp_41_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_41_cast_fp16 = tile(reps = v_exp_41_reps_0, x = var_2173_cast_fp16)[name = string("v_exp_41_cast_fp16")];
            tensor<int32, [4]> var_2176 = const()[name = string("op_2176"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_43_cast_fp16 = reshape(shape = var_2176, x = v_exp_41_cast_fp16)[name = string("v_exp_43_cast_fp16")];
            bool var_2179_transpose_x_1 = const()[name = string("op_2179_transpose_x_1"), val = bool(false)];
            bool var_2179_transpose_y_1 = const()[name = string("op_2179_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_2179_cast_fp16 = matmul(transpose_x = var_2179_transpose_x_1, transpose_y = var_2179_transpose_y_1, x = q_21_cast_fp16, y = k_exp_43_cast_fp16)[name = string("op_2179_cast_fp16")];
            fp16 var_2180_to_fp16 = const()[name = string("op_2180_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_41_cast_fp16 = mul(x = var_2179_cast_fp16, y = var_2180_to_fp16)[name = string("attn_41_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_101_cast_fp16 = add(x = attn_41_cast_fp16, y = attention_mask_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_43_cast_fp16 = softmax(axis = var_2037, x = input_101_cast_fp16)[name = string("attn_43_cast_fp16")];
            bool out_21_transpose_x_0 = const()[name = string("out_21_transpose_x_0"), val = bool(false)];
            bool out_21_transpose_y_0 = const()[name = string("out_21_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_43_cast_fp16, y = v_exp_43_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<int32, [4]> var_2185_perm_0 = const()[name = string("op_2185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2186 = const()[name = string("op_2186"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_2185_cast_fp16 = transpose(perm = var_2185_perm_0, x = out_21_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 128, 2048]> input_103_cast_fp16 = reshape(shape = var_2186, x = var_2185_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_10_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161591552))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163688768))))[name = string("layers_10_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_o_proj_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_279_cast_fp16 = add(x = x_259_cast_fp16, y = linear_73_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_2036_promoted_3_to_fp16 = const()[name = string("op_2036_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2193_cast_fp16 = pow(x = x_279_cast_fp16, y = var_2036_promoted_3_to_fp16)[name = string("op_2193_cast_fp16")];
            tensor<int32, [1]> var_2195_axes_0 = const()[name = string("op_2195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2195_keep_dims_0 = const()[name = string("op_2195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2195_cast_fp16 = reduce_mean(axes = var_2195_axes_0, keep_dims = var_2195_keep_dims_0, x = var_2193_cast_fp16)[name = string("op_2195_cast_fp16")];
            fp16 var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2197_cast_fp16 = add(x = var_2195_cast_fp16, y = var_2196_to_fp16)[name = string("op_2197_cast_fp16")];
            fp32 norm_87_epsilon_0 = const()[name = string("norm_87_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_87_cast_fp16 = rsqrt(epsilon = norm_87_epsilon_0, x = var_2197_cast_fp16)[name = string("norm_87_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2199_cast_fp16 = mul(x = x_279_cast_fp16, y = norm_87_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1024]> layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163689344)))];
            tensor<fp16, [1, 128, 1024]> var_2200_cast_fp16 = mul(x = var_2199_cast_fp16, y = layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_2200_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_10_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163691456))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166837248))))[name = string("layers_10_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_gate_proj_weight_to_fp16_palettized, x = var_2200_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_2210_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_10_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166837824))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169983616))))[name = string("layers_10_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_up_proj_weight_to_fp16_palettized, x = var_2200_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_109_cast_fp16 = mul(x = var_2210_cast_fp16, y = linear_75_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_10_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169984192))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173129984))))[name = string("layers_10_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_mlp_down_proj_weight_to_fp16_palettized, x = input_109_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_285_cast_fp16 = add(x = x_279_cast_fp16, y = linear_76_cast_fp16)[name = string("x_285_cast_fp16")];
            int32 var_2231 = const()[name = string("op_2231"), val = int32(-1)];
            fp16 var_2230_promoted_to_fp16 = const()[name = string("op_2230_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2240_cast_fp16 = pow(x = x_285_cast_fp16, y = var_2230_promoted_to_fp16)[name = string("op_2240_cast_fp16")];
            tensor<int32, [1]> var_2242_axes_0 = const()[name = string("op_2242_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2242_keep_dims_0 = const()[name = string("op_2242_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2242_cast_fp16 = reduce_mean(axes = var_2242_axes_0, keep_dims = var_2242_keep_dims_0, x = var_2240_cast_fp16)[name = string("op_2242_cast_fp16")];
            fp16 var_2243_to_fp16 = const()[name = string("op_2243_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2244_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_to_fp16)[name = string("op_2244_cast_fp16")];
            fp32 norm_89_epsilon_0 = const()[name = string("norm_89_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_89_cast_fp16 = rsqrt(epsilon = norm_89_epsilon_0, x = var_2244_cast_fp16)[name = string("norm_89_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2246_cast_fp16 = mul(x = x_285_cast_fp16, y = norm_89_cast_fp16)[name = string("op_2246_cast_fp16")];
            tensor<fp16, [1024]> layers_11_input_layernorm_weight_to_fp16 = const()[name = string("layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173130560)))];
            tensor<fp16, [1, 128, 1024]> var_2247_cast_fp16 = mul(x = var_2246_cast_fp16, y = layers_11_input_layernorm_weight_to_fp16)[name = string("op_2247_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_11_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173132672))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175229888))))[name = string("layers_11_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_q_proj_weight_to_fp16_palettized, x = var_2247_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_2263 = const()[name = string("op_2263"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_2264_cast_fp16 = reshape(shape = var_2263, x = linear_77_cast_fp16)[name = string("op_2264_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(175230464))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176279104))))[name = string("layers_11_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_k_proj_weight_to_fp16_palettized, x = var_2247_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_2268 = const()[name = string("op_2268"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2269_cast_fp16 = reshape(shape = var_2268, x = linear_78_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<int32, [4]> x_295_perm_0 = const()[name = string("x_295_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176279680))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177328320))))[name = string("layers_11_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_v_proj_weight_to_fp16_palettized, x = var_2247_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_2273 = const()[name = string("op_2273"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2274_cast_fp16 = reshape(shape = var_2273, x = linear_79_cast_fp16)[name = string("op_2274_cast_fp16")];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_2230_promoted_1_to_fp16 = const()[name = string("op_2230_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_2264_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 128, 128]> var_2278_cast_fp16 = pow(x = x_291_cast_fp16, y = var_2230_promoted_1_to_fp16)[name = string("op_2278_cast_fp16")];
            tensor<int32, [1]> var_2280_axes_0 = const()[name = string("op_2280_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2280_keep_dims_0 = const()[name = string("op_2280_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_2280_cast_fp16 = reduce_mean(axes = var_2280_axes_0, keep_dims = var_2280_keep_dims_0, x = var_2278_cast_fp16)[name = string("op_2280_cast_fp16")];
            fp16 var_2281_to_fp16 = const()[name = string("op_2281_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_2282_cast_fp16 = add(x = var_2280_cast_fp16, y = var_2281_to_fp16)[name = string("op_2282_cast_fp16")];
            fp32 norm_91_epsilon_0 = const()[name = string("norm_91_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_91_cast_fp16 = rsqrt(epsilon = norm_91_epsilon_0, x = var_2282_cast_fp16)[name = string("norm_91_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2284_cast_fp16 = mul(x = x_291_cast_fp16, y = norm_91_cast_fp16)[name = string("op_2284_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177328896)))];
            tensor<fp16, [1, 16, 128, 128]> var_2285_cast_fp16 = mul(x = var_2284_cast_fp16, y = layers_11_self_attn_q_norm_weight_to_fp16)[name = string("op_2285_cast_fp16")];
            fp16 var_2230_promoted_2_to_fp16 = const()[name = string("op_2230_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_295_cast_fp16 = transpose(perm = x_295_perm_0, x = var_2269_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 8, 128, 128]> var_2289_cast_fp16 = pow(x = x_295_cast_fp16, y = var_2230_promoted_2_to_fp16)[name = string("op_2289_cast_fp16")];
            tensor<int32, [1]> var_2291_axes_0 = const()[name = string("op_2291_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2291_keep_dims_0 = const()[name = string("op_2291_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_2291_cast_fp16 = reduce_mean(axes = var_2291_axes_0, keep_dims = var_2291_keep_dims_0, x = var_2289_cast_fp16)[name = string("op_2291_cast_fp16")];
            fp16 var_2292_to_fp16 = const()[name = string("op_2292_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_2293_cast_fp16 = add(x = var_2291_cast_fp16, y = var_2292_to_fp16)[name = string("op_2293_cast_fp16")];
            fp32 norm_93_epsilon_0 = const()[name = string("norm_93_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_93_cast_fp16 = rsqrt(epsilon = norm_93_epsilon_0, x = var_2293_cast_fp16)[name = string("norm_93_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_2295_cast_fp16 = mul(x = x_295_cast_fp16, y = norm_93_cast_fp16)[name = string("op_2295_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177329216)))];
            tensor<fp16, [1, 8, 128, 128]> var_2296_cast_fp16 = mul(x = var_2295_cast_fp16, y = layers_11_self_attn_k_norm_weight_to_fp16)[name = string("op_2296_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_2285_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, 64])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_2285_cast_fp16)[name = string("x2_45_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2317_cast_fp16 = mul(x = x1_45_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2318_cast_fp16 = mul(x = x2_45_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2319_cast_fp16 = sub(x = var_2317_cast_fp16, y = var_2318_cast_fp16)[name = string("op_2319_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2320_cast_fp16 = mul(x = x2_45_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2320_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2321_cast_fp16 = mul(x = x1_45_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2321_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2322_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2321_cast_fp16)[name = string("op_2322_cast_fp16")];
            bool q_23_interleave_0 = const()[name = string("q_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_23_cast_fp16 = concat(axis = var_2231, interleave = q_23_interleave_0, values = (var_2319_cast_fp16, var_2322_cast_fp16))[name = string("q_23_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_2296_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, 64])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_2296_cast_fp16)[name = string("x2_47_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2344_cast_fp16 = mul(x = x1_47_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2345_cast_fp16 = mul(x = x2_47_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2345_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2346_cast_fp16 = sub(x = var_2344_cast_fp16, y = var_2345_cast_fp16)[name = string("op_2346_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2347_cast_fp16 = mul(x = x2_47_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2347_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2348_cast_fp16 = mul(x = x1_47_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2348_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2349_cast_fp16 = add(x = var_2347_cast_fp16, y = var_2348_cast_fp16)[name = string("op_2349_cast_fp16")];
            bool var_2351_interleave_0 = const()[name = string("op_2351_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2351_cast_fp16 = concat(axis = var_2231, interleave = var_2351_interleave_0, values = (var_2346_cast_fp16, var_2349_cast_fp16))[name = string("op_2351_cast_fp16")];
            tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_202 = const()[name = string("concat_202"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_45_cast_fp16 = transpose(perm = transpose_45_perm_0, x = var_2351_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [128, 1024]> reshape_67_cast_fp16 = reshape(shape = concat_202, x = transpose_45_cast_fp16)[name = string("reshape_67_cast_fp16")];
            bool matmul_22_transpose_x_1 = const()[name = string("matmul_22_transpose_x_1"), val = bool(true)];
            bool matmul_22_transpose_y_1 = const()[name = string("matmul_22_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_22_cast_fp16 = matmul(transpose_x = matmul_22_transpose_x_1, transpose_y = matmul_22_transpose_y_1, x = var_66_to_fp16, y = reshape_67_cast_fp16)[name = string("matmul_22_cast_fp16")];
            tensor<int32, [4]> concat_205 = const()[name = string("concat_205"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_68_cast_fp16 = reshape(shape = concat_205, x = matmul_22_cast_fp16)[name = string("reshape_68_cast_fp16")];
            tensor<int32, [4]> scattered_k_23_perm_0 = const()[name = string("scattered_k_23_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_210 = const()[name = string("concat_210"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_67_cast_fp16 = transpose(perm = transpose_67_perm_0, x = var_2274_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [128, 1024]> reshape_70_cast_fp16 = reshape(shape = concat_210, x = transpose_67_cast_fp16)[name = string("reshape_70_cast_fp16")];
            bool matmul_23_transpose_x_1 = const()[name = string("matmul_23_transpose_x_1"), val = bool(true)];
            bool matmul_23_transpose_y_1 = const()[name = string("matmul_23_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_23_cast_fp16 = matmul(transpose_x = matmul_23_transpose_x_1, transpose_y = matmul_23_transpose_y_1, x = var_66_to_fp16, y = reshape_70_cast_fp16)[name = string("matmul_23_cast_fp16")];
            tensor<int32, [4]> concat_213 = const()[name = string("concat_213"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_71_cast_fp16 = reshape(shape = concat_213, x = matmul_23_cast_fp16)[name = string("reshape_71_cast_fp16")];
            tensor<int32, [4]> scattered_v_23_perm_0 = const()[name = string("scattered_v_23_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_22 = read_state(input = k_cache_11)[name = string("read_state_22")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_69_cast_fp16 = mul(x = read_state_22, y = var_222_cast_fp16)[name = string("k_cache_69_cast_fp16")];
            write_state(data = k_cache_69_cast_fp16, input = k_cache_11)[name = string("coreml_update_state_100_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_100 = read_state(input = k_cache_11)[name = string("coreml_update_state_100")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_23_cast_fp16 = transpose(perm = scattered_k_23_perm_0, x = reshape_68_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_71_cast_fp16 = add(x = coreml_update_state_100, y = scattered_k_23_cast_fp16)[name = string("k_cache_71_cast_fp16")];
            write_state(data = k_cache_71_cast_fp16, input = k_cache_11)[name = string("coreml_update_state_101_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_101 = read_state(input = k_cache_11)[name = string("coreml_update_state_101")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_23 = read_state(input = v_cache_11)[name = string("read_state_23")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_69_cast_fp16 = mul(x = read_state_23, y = var_222_cast_fp16)[name = string("v_cache_69_cast_fp16")];
            write_state(data = v_cache_69_cast_fp16, input = v_cache_11)[name = string("coreml_update_state_102_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_102 = read_state(input = v_cache_11)[name = string("coreml_update_state_102")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_23_cast_fp16 = transpose(perm = scattered_v_23_perm_0, x = reshape_71_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_71_cast_fp16 = add(x = coreml_update_state_102, y = scattered_v_23_cast_fp16)[name = string("v_cache_71_cast_fp16")];
            write_state(data = v_cache_71_cast_fp16, input = v_cache_11)[name = string("coreml_update_state_103_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_103 = read_state(input = v_cache_11)[name = string("coreml_update_state_103")];
            tensor<int32, [1]> var_2362_axes_0 = const()[name = string("op_2362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2362_cast_fp16 = expand_dims(axes = var_2362_axes_0, x = coreml_update_state_101)[name = string("op_2362_cast_fp16")];
            tensor<int32, [5]> k_exp_45_reps_0 = const()[name = string("k_exp_45_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_45_cast_fp16 = tile(reps = k_exp_45_reps_0, x = var_2362_cast_fp16)[name = string("k_exp_45_cast_fp16")];
            tensor<int32, [4]> var_2365 = const()[name = string("op_2365"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_47_cast_fp16 = reshape(shape = var_2365, x = k_exp_45_cast_fp16)[name = string("k_exp_47_cast_fp16")];
            tensor<int32, [1]> var_2367_axes_0 = const()[name = string("op_2367_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2367_cast_fp16 = expand_dims(axes = var_2367_axes_0, x = coreml_update_state_103)[name = string("op_2367_cast_fp16")];
            tensor<int32, [5]> v_exp_45_reps_0 = const()[name = string("v_exp_45_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_45_cast_fp16 = tile(reps = v_exp_45_reps_0, x = var_2367_cast_fp16)[name = string("v_exp_45_cast_fp16")];
            tensor<int32, [4]> var_2370 = const()[name = string("op_2370"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_47_cast_fp16 = reshape(shape = var_2370, x = v_exp_45_cast_fp16)[name = string("v_exp_47_cast_fp16")];
            bool var_2373_transpose_x_1 = const()[name = string("op_2373_transpose_x_1"), val = bool(false)];
            bool var_2373_transpose_y_1 = const()[name = string("op_2373_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_2373_cast_fp16 = matmul(transpose_x = var_2373_transpose_x_1, transpose_y = var_2373_transpose_y_1, x = q_23_cast_fp16, y = k_exp_47_cast_fp16)[name = string("op_2373_cast_fp16")];
            fp16 var_2374_to_fp16 = const()[name = string("op_2374_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_45_cast_fp16 = mul(x = var_2373_cast_fp16, y = var_2374_to_fp16)[name = string("attn_45_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_111_cast_fp16 = add(x = attn_45_cast_fp16, y = attention_mask_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_47_cast_fp16 = softmax(axis = var_2231, x = input_111_cast_fp16)[name = string("attn_47_cast_fp16")];
            bool out_23_transpose_x_0 = const()[name = string("out_23_transpose_x_0"), val = bool(false)];
            bool out_23_transpose_y_0 = const()[name = string("out_23_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_23_cast_fp16 = matmul(transpose_x = out_23_transpose_x_0, transpose_y = out_23_transpose_y_0, x = attn_47_cast_fp16, y = v_exp_47_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<int32, [4]> var_2379_perm_0 = const()[name = string("op_2379_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2380 = const()[name = string("op_2380"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_2379_cast_fp16 = transpose(perm = var_2379_perm_0, x = out_23_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 128, 2048]> input_113_cast_fp16 = reshape(shape = var_2380, x = var_2379_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_11_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177329536))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179426752))))[name = string("layers_11_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_o_proj_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_305_cast_fp16 = add(x = x_285_cast_fp16, y = linear_80_cast_fp16)[name = string("x_305_cast_fp16")];
            fp16 var_2230_promoted_3_to_fp16 = const()[name = string("op_2230_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2387_cast_fp16 = pow(x = x_305_cast_fp16, y = var_2230_promoted_3_to_fp16)[name = string("op_2387_cast_fp16")];
            tensor<int32, [1]> var_2389_axes_0 = const()[name = string("op_2389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2389_keep_dims_0 = const()[name = string("op_2389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2389_cast_fp16 = reduce_mean(axes = var_2389_axes_0, keep_dims = var_2389_keep_dims_0, x = var_2387_cast_fp16)[name = string("op_2389_cast_fp16")];
            fp16 var_2390_to_fp16 = const()[name = string("op_2390_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2391_cast_fp16 = add(x = var_2389_cast_fp16, y = var_2390_to_fp16)[name = string("op_2391_cast_fp16")];
            fp32 norm_95_epsilon_0 = const()[name = string("norm_95_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_95_cast_fp16 = rsqrt(epsilon = norm_95_epsilon_0, x = var_2391_cast_fp16)[name = string("norm_95_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2393_cast_fp16 = mul(x = x_305_cast_fp16, y = norm_95_cast_fp16)[name = string("op_2393_cast_fp16")];
            tensor<fp16, [1024]> layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179427328)))];
            tensor<fp16, [1, 128, 1024]> var_2394_cast_fp16 = mul(x = var_2393_cast_fp16, y = layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_2394_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_11_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179429440))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182575232))))[name = string("layers_11_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_gate_proj_weight_to_fp16_palettized, x = var_2394_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_2404_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_11_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182575808))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185721600))))[name = string("layers_11_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_up_proj_weight_to_fp16_palettized, x = var_2394_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_119_cast_fp16 = mul(x = var_2404_cast_fp16, y = linear_82_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_11_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185722176))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188867968))))[name = string("layers_11_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_mlp_down_proj_weight_to_fp16_palettized, x = input_119_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_311_cast_fp16 = add(x = x_305_cast_fp16, y = linear_83_cast_fp16)[name = string("x_311_cast_fp16")];
            int32 var_2425 = const()[name = string("op_2425"), val = int32(-1)];
            fp16 var_2424_promoted_to_fp16 = const()[name = string("op_2424_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2434_cast_fp16 = pow(x = x_311_cast_fp16, y = var_2424_promoted_to_fp16)[name = string("op_2434_cast_fp16")];
            tensor<int32, [1]> var_2436_axes_0 = const()[name = string("op_2436_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2436_keep_dims_0 = const()[name = string("op_2436_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2436_cast_fp16 = reduce_mean(axes = var_2436_axes_0, keep_dims = var_2436_keep_dims_0, x = var_2434_cast_fp16)[name = string("op_2436_cast_fp16")];
            fp16 var_2437_to_fp16 = const()[name = string("op_2437_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2438_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2437_to_fp16)[name = string("op_2438_cast_fp16")];
            fp32 norm_97_epsilon_0 = const()[name = string("norm_97_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_97_cast_fp16 = rsqrt(epsilon = norm_97_epsilon_0, x = var_2438_cast_fp16)[name = string("norm_97_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2440_cast_fp16 = mul(x = x_311_cast_fp16, y = norm_97_cast_fp16)[name = string("op_2440_cast_fp16")];
            tensor<fp16, [1024]> layers_12_input_layernorm_weight_to_fp16 = const()[name = string("layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188868544)))];
            tensor<fp16, [1, 128, 1024]> var_2441_cast_fp16 = mul(x = var_2440_cast_fp16, y = layers_12_input_layernorm_weight_to_fp16)[name = string("op_2441_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_12_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188870656))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190967872))))[name = string("layers_12_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_q_proj_weight_to_fp16_palettized, x = var_2441_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_2457 = const()[name = string("op_2457"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_2458_cast_fp16 = reshape(shape = var_2457, x = linear_84_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<int32, [4]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190968448))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192017088))))[name = string("layers_12_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_k_proj_weight_to_fp16_palettized, x = var_2441_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_2462 = const()[name = string("op_2462"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2463_cast_fp16 = reshape(shape = var_2462, x = linear_85_cast_fp16)[name = string("op_2463_cast_fp16")];
            tensor<int32, [4]> x_321_perm_0 = const()[name = string("x_321_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192017664))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193066304))))[name = string("layers_12_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_v_proj_weight_to_fp16_palettized, x = var_2441_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2468_cast_fp16 = reshape(shape = var_2467, x = linear_86_cast_fp16)[name = string("op_2468_cast_fp16")];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_2424_promoted_1_to_fp16 = const()[name = string("op_2424_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = var_2458_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 16, 128, 128]> var_2472_cast_fp16 = pow(x = x_317_cast_fp16, y = var_2424_promoted_1_to_fp16)[name = string("op_2472_cast_fp16")];
            tensor<int32, [1]> var_2474_axes_0 = const()[name = string("op_2474_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2474_keep_dims_0 = const()[name = string("op_2474_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_2474_cast_fp16 = reduce_mean(axes = var_2474_axes_0, keep_dims = var_2474_keep_dims_0, x = var_2472_cast_fp16)[name = string("op_2474_cast_fp16")];
            fp16 var_2475_to_fp16 = const()[name = string("op_2475_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_2476_cast_fp16 = add(x = var_2474_cast_fp16, y = var_2475_to_fp16)[name = string("op_2476_cast_fp16")];
            fp32 norm_99_epsilon_0 = const()[name = string("norm_99_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_99_cast_fp16 = rsqrt(epsilon = norm_99_epsilon_0, x = var_2476_cast_fp16)[name = string("norm_99_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2478_cast_fp16 = mul(x = x_317_cast_fp16, y = norm_99_cast_fp16)[name = string("op_2478_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193066880)))];
            tensor<fp16, [1, 16, 128, 128]> var_2479_cast_fp16 = mul(x = var_2478_cast_fp16, y = layers_12_self_attn_q_norm_weight_to_fp16)[name = string("op_2479_cast_fp16")];
            fp16 var_2424_promoted_2_to_fp16 = const()[name = string("op_2424_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_321_cast_fp16 = transpose(perm = x_321_perm_0, x = var_2463_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 128, 128]> var_2483_cast_fp16 = pow(x = x_321_cast_fp16, y = var_2424_promoted_2_to_fp16)[name = string("op_2483_cast_fp16")];
            tensor<int32, [1]> var_2485_axes_0 = const()[name = string("op_2485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2485_keep_dims_0 = const()[name = string("op_2485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_2485_cast_fp16 = reduce_mean(axes = var_2485_axes_0, keep_dims = var_2485_keep_dims_0, x = var_2483_cast_fp16)[name = string("op_2485_cast_fp16")];
            fp16 var_2486_to_fp16 = const()[name = string("op_2486_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_2487_cast_fp16 = add(x = var_2485_cast_fp16, y = var_2486_to_fp16)[name = string("op_2487_cast_fp16")];
            fp32 norm_101_epsilon_0 = const()[name = string("norm_101_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_101_cast_fp16 = rsqrt(epsilon = norm_101_epsilon_0, x = var_2487_cast_fp16)[name = string("norm_101_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_2489_cast_fp16 = mul(x = x_321_cast_fp16, y = norm_101_cast_fp16)[name = string("op_2489_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193067200)))];
            tensor<fp16, [1, 8, 128, 128]> var_2490_cast_fp16 = mul(x = var_2489_cast_fp16, y = layers_12_self_attn_k_norm_weight_to_fp16)[name = string("op_2490_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_2479_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, 64])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_2479_cast_fp16)[name = string("x2_49_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2511_cast_fp16 = mul(x = x1_49_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2512_cast_fp16 = mul(x = x2_49_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2512_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2513_cast_fp16 = sub(x = var_2511_cast_fp16, y = var_2512_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2514_cast_fp16 = mul(x = x2_49_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2514_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2515_cast_fp16 = mul(x = x1_49_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2515_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2516_cast_fp16 = add(x = var_2514_cast_fp16, y = var_2515_cast_fp16)[name = string("op_2516_cast_fp16")];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_25_cast_fp16 = concat(axis = var_2425, interleave = q_25_interleave_0, values = (var_2513_cast_fp16, var_2516_cast_fp16))[name = string("q_25_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, 8, 128, 64])];
            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, 8, 128, 64]> 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 = var_2490_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, 64])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> 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 = var_2490_cast_fp16)[name = string("x2_51_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2538_cast_fp16 = mul(x = x1_51_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2538_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2539_cast_fp16 = mul(x = x2_51_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2539_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2540_cast_fp16 = sub(x = var_2538_cast_fp16, y = var_2539_cast_fp16)[name = string("op_2540_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2541_cast_fp16 = mul(x = x2_51_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2541_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2542_cast_fp16 = mul(x = x1_51_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2542_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2543_cast_fp16 = add(x = var_2541_cast_fp16, y = var_2542_cast_fp16)[name = string("op_2543_cast_fp16")];
            bool var_2545_interleave_0 = const()[name = string("op_2545_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2545_cast_fp16 = concat(axis = var_2425, interleave = var_2545_interleave_0, values = (var_2540_cast_fp16, var_2543_cast_fp16))[name = string("op_2545_cast_fp16")];
            tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_220 = const()[name = string("concat_220"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_49_cast_fp16 = transpose(perm = transpose_49_perm_0, x = var_2545_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [128, 1024]> reshape_73_cast_fp16 = reshape(shape = concat_220, x = transpose_49_cast_fp16)[name = string("reshape_73_cast_fp16")];
            bool matmul_24_transpose_x_1 = const()[name = string("matmul_24_transpose_x_1"), val = bool(true)];
            bool matmul_24_transpose_y_1 = const()[name = string("matmul_24_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_24_cast_fp16 = matmul(transpose_x = matmul_24_transpose_x_1, transpose_y = matmul_24_transpose_y_1, x = var_66_to_fp16, y = reshape_73_cast_fp16)[name = string("matmul_24_cast_fp16")];
            tensor<int32, [4]> concat_223 = const()[name = string("concat_223"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_74_cast_fp16 = reshape(shape = concat_223, x = matmul_24_cast_fp16)[name = string("reshape_74_cast_fp16")];
            tensor<int32, [4]> scattered_k_25_perm_0 = const()[name = string("scattered_k_25_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_228 = const()[name = string("concat_228"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_68_cast_fp16 = transpose(perm = transpose_68_perm_0, x = var_2468_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [128, 1024]> reshape_76_cast_fp16 = reshape(shape = concat_228, x = transpose_68_cast_fp16)[name = string("reshape_76_cast_fp16")];
            bool matmul_25_transpose_x_1 = const()[name = string("matmul_25_transpose_x_1"), val = bool(true)];
            bool matmul_25_transpose_y_1 = const()[name = string("matmul_25_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_25_cast_fp16 = matmul(transpose_x = matmul_25_transpose_x_1, transpose_y = matmul_25_transpose_y_1, x = var_66_to_fp16, y = reshape_76_cast_fp16)[name = string("matmul_25_cast_fp16")];
            tensor<int32, [4]> concat_231 = const()[name = string("concat_231"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_77_cast_fp16 = reshape(shape = concat_231, x = matmul_25_cast_fp16)[name = string("reshape_77_cast_fp16")];
            tensor<int32, [4]> scattered_v_25_perm_0 = const()[name = string("scattered_v_25_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_24 = read_state(input = k_cache_12)[name = string("read_state_24")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_75_cast_fp16 = mul(x = read_state_24, y = var_222_cast_fp16)[name = string("k_cache_75_cast_fp16")];
            write_state(data = k_cache_75_cast_fp16, input = k_cache_12)[name = string("coreml_update_state_104_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_104 = read_state(input = k_cache_12)[name = string("coreml_update_state_104")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_25_cast_fp16 = transpose(perm = scattered_k_25_perm_0, x = reshape_74_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_77_cast_fp16 = add(x = coreml_update_state_104, y = scattered_k_25_cast_fp16)[name = string("k_cache_77_cast_fp16")];
            write_state(data = k_cache_77_cast_fp16, input = k_cache_12)[name = string("coreml_update_state_105_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_105 = read_state(input = k_cache_12)[name = string("coreml_update_state_105")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_25 = read_state(input = v_cache_12)[name = string("read_state_25")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_75_cast_fp16 = mul(x = read_state_25, y = var_222_cast_fp16)[name = string("v_cache_75_cast_fp16")];
            write_state(data = v_cache_75_cast_fp16, input = v_cache_12)[name = string("coreml_update_state_106_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_106 = read_state(input = v_cache_12)[name = string("coreml_update_state_106")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_25_cast_fp16 = transpose(perm = scattered_v_25_perm_0, x = reshape_77_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_77_cast_fp16 = add(x = coreml_update_state_106, y = scattered_v_25_cast_fp16)[name = string("v_cache_77_cast_fp16")];
            write_state(data = v_cache_77_cast_fp16, input = v_cache_12)[name = string("coreml_update_state_107_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_107 = read_state(input = v_cache_12)[name = string("coreml_update_state_107")];
            tensor<int32, [1]> var_2556_axes_0 = const()[name = string("op_2556_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2556_cast_fp16 = expand_dims(axes = var_2556_axes_0, x = coreml_update_state_105)[name = string("op_2556_cast_fp16")];
            tensor<int32, [5]> k_exp_49_reps_0 = const()[name = string("k_exp_49_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_49_cast_fp16 = tile(reps = k_exp_49_reps_0, x = var_2556_cast_fp16)[name = string("k_exp_49_cast_fp16")];
            tensor<int32, [4]> var_2559 = const()[name = string("op_2559"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_51_cast_fp16 = reshape(shape = var_2559, x = k_exp_49_cast_fp16)[name = string("k_exp_51_cast_fp16")];
            tensor<int32, [1]> var_2561_axes_0 = const()[name = string("op_2561_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2561_cast_fp16 = expand_dims(axes = var_2561_axes_0, x = coreml_update_state_107)[name = string("op_2561_cast_fp16")];
            tensor<int32, [5]> v_exp_49_reps_0 = const()[name = string("v_exp_49_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_49_cast_fp16 = tile(reps = v_exp_49_reps_0, x = var_2561_cast_fp16)[name = string("v_exp_49_cast_fp16")];
            tensor<int32, [4]> var_2564 = const()[name = string("op_2564"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_51_cast_fp16 = reshape(shape = var_2564, x = v_exp_49_cast_fp16)[name = string("v_exp_51_cast_fp16")];
            bool var_2567_transpose_x_1 = const()[name = string("op_2567_transpose_x_1"), val = bool(false)];
            bool var_2567_transpose_y_1 = const()[name = string("op_2567_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_2567_cast_fp16 = matmul(transpose_x = var_2567_transpose_x_1, transpose_y = var_2567_transpose_y_1, x = q_25_cast_fp16, y = k_exp_51_cast_fp16)[name = string("op_2567_cast_fp16")];
            fp16 var_2568_to_fp16 = const()[name = string("op_2568_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_49_cast_fp16 = mul(x = var_2567_cast_fp16, y = var_2568_to_fp16)[name = string("attn_49_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_121_cast_fp16 = add(x = attn_49_cast_fp16, y = attention_mask_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_51_cast_fp16 = softmax(axis = var_2425, x = input_121_cast_fp16)[name = string("attn_51_cast_fp16")];
            bool out_25_transpose_x_0 = const()[name = string("out_25_transpose_x_0"), val = bool(false)];
            bool out_25_transpose_y_0 = const()[name = string("out_25_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = attn_51_cast_fp16, y = v_exp_51_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<int32, [4]> var_2573_perm_0 = const()[name = string("op_2573_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2574 = const()[name = string("op_2574"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_2573_cast_fp16 = transpose(perm = var_2573_perm_0, x = out_25_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 128, 2048]> input_123_cast_fp16 = reshape(shape = var_2574, x = var_2573_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_12_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193067520))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195164736))))[name = string("layers_12_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_o_proj_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_331_cast_fp16 = add(x = x_311_cast_fp16, y = linear_87_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_2424_promoted_3_to_fp16 = const()[name = string("op_2424_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2581_cast_fp16 = pow(x = x_331_cast_fp16, y = var_2424_promoted_3_to_fp16)[name = string("op_2581_cast_fp16")];
            tensor<int32, [1]> var_2583_axes_0 = const()[name = string("op_2583_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2583_keep_dims_0 = const()[name = string("op_2583_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2583_cast_fp16 = reduce_mean(axes = var_2583_axes_0, keep_dims = var_2583_keep_dims_0, x = var_2581_cast_fp16)[name = string("op_2583_cast_fp16")];
            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2585_cast_fp16 = add(x = var_2583_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
            fp32 norm_103_epsilon_0 = const()[name = string("norm_103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_103_cast_fp16 = rsqrt(epsilon = norm_103_epsilon_0, x = var_2585_cast_fp16)[name = string("norm_103_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2587_cast_fp16 = mul(x = x_331_cast_fp16, y = norm_103_cast_fp16)[name = string("op_2587_cast_fp16")];
            tensor<fp16, [1024]> layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195165312)))];
            tensor<fp16, [1, 128, 1024]> var_2588_cast_fp16 = mul(x = var_2587_cast_fp16, y = layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_12_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195167424))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198313216))))[name = string("layers_12_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_gate_proj_weight_to_fp16_palettized, x = var_2588_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_2598_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_2598_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_12_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198313792))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201459584))))[name = string("layers_12_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_up_proj_weight_to_fp16_palettized, x = var_2588_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_129_cast_fp16 = mul(x = var_2598_cast_fp16, y = linear_89_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_12_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201460160))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204605952))))[name = string("layers_12_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_mlp_down_proj_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_337_cast_fp16 = add(x = x_331_cast_fp16, y = linear_90_cast_fp16)[name = string("x_337_cast_fp16")];
            int32 var_2619 = const()[name = string("op_2619"), val = int32(-1)];
            fp16 var_2618_promoted_to_fp16 = const()[name = string("op_2618_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2628_cast_fp16 = pow(x = x_337_cast_fp16, y = var_2618_promoted_to_fp16)[name = string("op_2628_cast_fp16")];
            tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2630_keep_dims_0 = const()[name = string("op_2630_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2630_cast_fp16 = reduce_mean(axes = var_2630_axes_0, keep_dims = var_2630_keep_dims_0, x = var_2628_cast_fp16)[name = string("op_2630_cast_fp16")];
            fp16 var_2631_to_fp16 = const()[name = string("op_2631_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2632_cast_fp16 = add(x = var_2630_cast_fp16, y = var_2631_to_fp16)[name = string("op_2632_cast_fp16")];
            fp32 norm_105_epsilon_0 = const()[name = string("norm_105_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_105_cast_fp16 = rsqrt(epsilon = norm_105_epsilon_0, x = var_2632_cast_fp16)[name = string("norm_105_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2634_cast_fp16 = mul(x = x_337_cast_fp16, y = norm_105_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1024]> layers_13_input_layernorm_weight_to_fp16 = const()[name = string("layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204606528)))];
            tensor<fp16, [1, 128, 1024]> var_2635_cast_fp16 = mul(x = var_2634_cast_fp16, y = layers_13_input_layernorm_weight_to_fp16)[name = string("op_2635_cast_fp16")];
            tensor<fp16, [2048, 1024]> layers_13_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(204608640))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206705856))))[name = string("layers_13_self_attn_q_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_q_proj_weight_to_fp16_palettized, x = var_2635_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2651 = const()[name = string("op_2651"), val = tensor<int32, [4]>([1, 128, 16, 128])];
            tensor<fp16, [1, 128, 16, 128]> var_2652_cast_fp16 = reshape(shape = var_2651, x = linear_91_cast_fp16)[name = string("op_2652_cast_fp16")];
            tensor<int32, [4]> x_343_perm_0 = const()[name = string("x_343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206706432))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207755072))))[name = string("layers_13_self_attn_k_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_k_proj_weight_to_fp16_palettized, x = var_2635_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2656 = const()[name = string("op_2656"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2657_cast_fp16 = reshape(shape = var_2656, x = linear_92_cast_fp16)[name = string("op_2657_cast_fp16")];
            tensor<int32, [4]> x_347_perm_0 = const()[name = string("x_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207755648))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208804288))))[name = string("layers_13_self_attn_v_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_v_proj_weight_to_fp16_palettized, x = var_2635_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2661 = const()[name = string("op_2661"), val = tensor<int32, [4]>([1, 128, 8, 128])];
            tensor<fp16, [1, 128, 8, 128]> var_2662_cast_fp16 = reshape(shape = var_2661, x = linear_93_cast_fp16)[name = string("op_2662_cast_fp16")];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
            fp16 var_2618_promoted_1_to_fp16 = const()[name = string("op_2618_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 128, 128]> x_343_cast_fp16 = transpose(perm = x_343_perm_0, x = var_2652_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 16, 128, 128]> var_2666_cast_fp16 = pow(x = x_343_cast_fp16, y = var_2618_promoted_1_to_fp16)[name = string("op_2666_cast_fp16")];
            tensor<int32, [1]> var_2668_axes_0 = const()[name = string("op_2668_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2668_keep_dims_0 = const()[name = string("op_2668_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1]> var_2668_cast_fp16 = reduce_mean(axes = var_2668_axes_0, keep_dims = var_2668_keep_dims_0, x = var_2666_cast_fp16)[name = string("op_2668_cast_fp16")];
            fp16 var_2669_to_fp16 = const()[name = string("op_2669_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 128, 1]> var_2670_cast_fp16 = add(x = var_2668_cast_fp16, y = var_2669_to_fp16)[name = string("op_2670_cast_fp16")];
            fp32 norm_107_epsilon_0 = const()[name = string("norm_107_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 128, 1]> norm_107_cast_fp16 = rsqrt(epsilon = norm_107_epsilon_0, x = var_2670_cast_fp16)[name = string("norm_107_cast_fp16")];
            tensor<fp16, [1, 16, 128, 128]> var_2672_cast_fp16 = mul(x = x_343_cast_fp16, y = norm_107_cast_fp16)[name = string("op_2672_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208804864)))];
            tensor<fp16, [1, 16, 128, 128]> var_2673_cast_fp16 = mul(x = var_2672_cast_fp16, y = layers_13_self_attn_q_norm_weight_to_fp16)[name = string("op_2673_cast_fp16")];
            fp16 var_2618_promoted_2_to_fp16 = const()[name = string("op_2618_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 128, 128]> x_347_cast_fp16 = transpose(perm = x_347_perm_0, x = var_2657_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 8, 128, 128]> var_2677_cast_fp16 = pow(x = x_347_cast_fp16, y = var_2618_promoted_2_to_fp16)[name = string("op_2677_cast_fp16")];
            tensor<int32, [1]> var_2679_axes_0 = const()[name = string("op_2679_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2679_keep_dims_0 = const()[name = string("op_2679_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 128, 1]> var_2679_cast_fp16 = reduce_mean(axes = var_2679_axes_0, keep_dims = var_2679_keep_dims_0, x = var_2677_cast_fp16)[name = string("op_2679_cast_fp16")];
            fp16 var_2680_to_fp16 = const()[name = string("op_2680_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 128, 1]> var_2681_cast_fp16 = add(x = var_2679_cast_fp16, y = var_2680_to_fp16)[name = string("op_2681_cast_fp16")];
            fp32 norm_109_epsilon_0 = const()[name = string("norm_109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 128, 1]> norm_109_cast_fp16 = rsqrt(epsilon = norm_109_epsilon_0, x = var_2681_cast_fp16)[name = string("norm_109_cast_fp16")];
            tensor<fp16, [1, 8, 128, 128]> var_2683_cast_fp16 = mul(x = x_347_cast_fp16, y = norm_109_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208805184)))];
            tensor<fp16, [1, 8, 128, 128]> var_2684_cast_fp16 = mul(x = var_2683_cast_fp16, y = layers_13_self_attn_k_norm_weight_to_fp16)[name = string("op_2684_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, 16, 128, 64])];
            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, 16, 128, 64]> 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 = var_2673_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, 64])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 128, 128])];
            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, 16, 128, 64]> 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 = var_2673_cast_fp16)[name = string("x2_53_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2705_cast_fp16 = mul(x = x1_53_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2706_cast_fp16 = mul(x = x2_53_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2706_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2707_cast_fp16 = sub(x = var_2705_cast_fp16, y = var_2706_cast_fp16)[name = string("op_2707_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2708_cast_fp16 = mul(x = x2_53_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2708_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2709_cast_fp16 = mul(x = x1_53_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2709_cast_fp16")];
            tensor<fp16, [1, 16, 128, 64]> var_2710_cast_fp16 = add(x = var_2708_cast_fp16, y = var_2709_cast_fp16)[name = string("op_2710_cast_fp16")];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> q_cast_fp16 = concat(axis = var_2619, interleave = q_interleave_0, values = (var_2707_cast_fp16, var_2710_cast_fp16))[name = string("q_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, 8, 128, 64])];
            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, 8, 128, 64]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = var_2684_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, 64])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 8, 128, 128])];
            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, 8, 128, 64]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = var_2684_cast_fp16)[name = string("x2_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2732_cast_fp16 = mul(x = x1_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2732_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2733_cast_fp16 = mul(x = x2_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2733_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2734_cast_fp16 = sub(x = var_2732_cast_fp16, y = var_2733_cast_fp16)[name = string("op_2734_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2735_cast_fp16 = mul(x = x2_cast_fp16, y = cos_val_1_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2736_cast_fp16 = mul(x = x1_cast_fp16, y = sin_val_1_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<fp16, [1, 8, 128, 64]> var_2737_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2736_cast_fp16)[name = string("op_2737_cast_fp16")];
            bool var_2739_interleave_0 = const()[name = string("op_2739_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 128, 128]> var_2739_cast_fp16 = concat(axis = var_2619, interleave = var_2739_interleave_0, values = (var_2734_cast_fp16, var_2737_cast_fp16))[name = string("op_2739_cast_fp16")];
            tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([2, 0, 1, 3])];
            tensor<int32, [2]> concat_238 = const()[name = string("concat_238"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_53_cast_fp16 = transpose(perm = transpose_53_perm_0, x = var_2739_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [128, 1024]> reshape_79_cast_fp16 = reshape(shape = concat_238, x = transpose_53_cast_fp16)[name = string("reshape_79_cast_fp16")];
            bool matmul_26_transpose_x_1 = const()[name = string("matmul_26_transpose_x_1"), val = bool(true)];
            bool matmul_26_transpose_y_1 = const()[name = string("matmul_26_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_26_cast_fp16 = matmul(transpose_x = matmul_26_transpose_x_1, transpose_y = matmul_26_transpose_y_1, x = var_66_to_fp16, y = reshape_79_cast_fp16)[name = string("matmul_26_cast_fp16")];
            tensor<int32, [4]> concat_241 = const()[name = string("concat_241"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_80_cast_fp16 = reshape(shape = concat_241, x = matmul_26_cast_fp16)[name = string("reshape_80_cast_fp16")];
            tensor<int32, [4]> scattered_k_perm_0 = const()[name = string("scattered_k_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<int32, [2]> concat_246 = const()[name = string("concat_246"), val = tensor<int32, [2]>([128, 1024])];
            tensor<fp16, [128, 1, 8, 128]> transpose_69_cast_fp16 = transpose(perm = transpose_69_perm_0, x = var_2662_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [128, 1024]> reshape_82_cast_fp16 = reshape(shape = concat_246, x = transpose_69_cast_fp16)[name = string("reshape_82_cast_fp16")];
            bool matmul_27_transpose_x_1 = const()[name = string("matmul_27_transpose_x_1"), val = bool(true)];
            bool matmul_27_transpose_y_1 = const()[name = string("matmul_27_transpose_y_1"), val = bool(false)];
            tensor<fp16, [1024, 1024]> matmul_27_cast_fp16 = matmul(transpose_x = matmul_27_transpose_x_1, transpose_y = matmul_27_transpose_y_1, x = var_66_to_fp16, y = reshape_82_cast_fp16)[name = string("matmul_27_cast_fp16")];
            tensor<int32, [4]> concat_249 = const()[name = string("concat_249"), val = tensor<int32, [4]>([1024, 1, 8, 128])];
            tensor<fp16, [1024, 1, 8, 128]> reshape_83_cast_fp16 = reshape(shape = concat_249, x = matmul_27_cast_fp16)[name = string("reshape_83_cast_fp16")];
            tensor<int32, [4]> scattered_v_perm_0 = const()[name = string("scattered_v_perm_0"), val = tensor<int32, [4]>([1, 2, 0, 3])];
            tensor<fp16, [1, 8, 1024, 128]> read_state_26 = read_state(input = k_cache_13)[name = string("read_state_26")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_81_cast_fp16 = mul(x = read_state_26, y = var_222_cast_fp16)[name = string("k_cache_81_cast_fp16")];
            write_state(data = k_cache_81_cast_fp16, input = k_cache_13)[name = string("coreml_update_state_108_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_108 = read_state(input = k_cache_13)[name = string("coreml_update_state_108")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_k_cast_fp16 = transpose(perm = scattered_k_perm_0, x = reshape_80_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 1024, 128]> k_cache_cast_fp16 = add(x = coreml_update_state_108, y = scattered_k_cast_fp16)[name = string("k_cache_cast_fp16")];
            write_state(data = k_cache_cast_fp16, input = k_cache_13)[name = string("coreml_update_state_109_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_109 = read_state(input = k_cache_13)[name = string("coreml_update_state_109")];
            tensor<fp16, [1, 8, 1024, 128]> read_state_27 = read_state(input = v_cache_13)[name = string("read_state_27")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_81_cast_fp16 = mul(x = read_state_27, y = var_222_cast_fp16)[name = string("v_cache_81_cast_fp16")];
            write_state(data = v_cache_81_cast_fp16, input = v_cache_13)[name = string("coreml_update_state_110_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_110 = read_state(input = v_cache_13)[name = string("coreml_update_state_110")];
            tensor<fp16, [1, 8, 1024, 128]> scattered_v_cast_fp16 = transpose(perm = scattered_v_perm_0, x = reshape_83_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 1024, 128]> v_cache_cast_fp16 = add(x = coreml_update_state_110, y = scattered_v_cast_fp16)[name = string("v_cache_cast_fp16")];
            write_state(data = v_cache_cast_fp16, input = v_cache_13)[name = string("coreml_update_state_111_write_state")];
            tensor<fp16, [1, 8, 1024, 128]> coreml_update_state_111 = read_state(input = v_cache_13)[name = string("coreml_update_state_111")];
            tensor<int32, [1]> var_2750_axes_0 = const()[name = string("op_2750_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2750_cast_fp16 = expand_dims(axes = var_2750_axes_0, x = coreml_update_state_109)[name = string("op_2750_cast_fp16")];
            tensor<int32, [5]> k_exp_53_reps_0 = const()[name = string("k_exp_53_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> k_exp_53_cast_fp16 = tile(reps = k_exp_53_reps_0, x = var_2750_cast_fp16)[name = string("k_exp_53_cast_fp16")];
            tensor<int32, [4]> var_2753 = const()[name = string("op_2753"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> k_exp_cast_fp16 = reshape(shape = var_2753, x = k_exp_53_cast_fp16)[name = string("k_exp_cast_fp16")];
            tensor<int32, [1]> var_2755_axes_0 = const()[name = string("op_2755_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 1024, 128]> var_2755_cast_fp16 = expand_dims(axes = var_2755_axes_0, x = coreml_update_state_111)[name = string("op_2755_cast_fp16")];
            tensor<int32, [5]> v_exp_53_reps_0 = const()[name = string("v_exp_53_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 1024, 128]> v_exp_53_cast_fp16 = tile(reps = v_exp_53_reps_0, x = var_2755_cast_fp16)[name = string("v_exp_53_cast_fp16")];
            tensor<int32, [4]> var_2758 = const()[name = string("op_2758"), val = tensor<int32, [4]>([1, 16, 1024, 128])];
            tensor<fp16, [1, 16, 1024, 128]> v_exp_cast_fp16 = reshape(shape = var_2758, x = v_exp_53_cast_fp16)[name = string("v_exp_cast_fp16")];
            bool var_2761_transpose_x_1 = const()[name = string("op_2761_transpose_x_1"), val = bool(false)];
            bool var_2761_transpose_y_1 = const()[name = string("op_2761_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 128, 1024]> var_2761_cast_fp16 = matmul(transpose_x = var_2761_transpose_x_1, transpose_y = var_2761_transpose_y_1, x = q_cast_fp16, y = k_exp_cast_fp16)[name = string("op_2761_cast_fp16")];
            fp16 var_2762_to_fp16 = const()[name = string("op_2762_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 128, 1024]> attn_53_cast_fp16 = mul(x = var_2761_cast_fp16, y = var_2762_to_fp16)[name = string("attn_53_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> input_131_cast_fp16 = add(x = attn_53_cast_fp16, y = attention_mask_to_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1, 16, 128, 1024]> attn_cast_fp16 = softmax(axis = var_2619, x = input_131_cast_fp16)[name = string("attn_cast_fp16")];
            bool out_transpose_x_0 = const()[name = string("out_transpose_x_0"), val = bool(false)];
            bool out_transpose_y_0 = const()[name = string("out_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 128, 128]> out_cast_fp16 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_cast_fp16, y = v_exp_cast_fp16)[name = string("out_cast_fp16")];
            tensor<int32, [4]> var_2767_perm_0 = const()[name = string("op_2767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2768 = const()[name = string("op_2768"), val = tensor<int32, [3]>([1, 128, -1])];
            tensor<fp16, [1, 128, 16, 128]> var_2767_cast_fp16 = transpose(perm = var_2767_perm_0, x = out_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 128, 2048]> input_133_cast_fp16 = reshape(shape = var_2768, x = var_2767_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 2048]> layers_13_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208805504))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210902720))))[name = string("layers_13_self_attn_o_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_o_proj_weight_to_fp16_palettized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> x_357_cast_fp16 = add(x = x_337_cast_fp16, y = linear_94_cast_fp16)[name = string("x_357_cast_fp16")];
            fp16 var_2618_promoted_3_to_fp16 = const()[name = string("op_2618_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 128, 1024]> var_2775_cast_fp16 = pow(x = x_357_cast_fp16, y = var_2618_promoted_3_to_fp16)[name = string("op_2775_cast_fp16")];
            tensor<int32, [1]> var_2777_axes_0 = const()[name = string("op_2777_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2777_keep_dims_0 = const()[name = string("op_2777_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 128, 1]> var_2777_cast_fp16 = reduce_mean(axes = var_2777_axes_0, keep_dims = var_2777_keep_dims_0, x = var_2775_cast_fp16)[name = string("op_2777_cast_fp16")];
            fp16 var_2778_to_fp16 = const()[name = string("op_2778_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 128, 1]> var_2779_cast_fp16 = add(x = var_2777_cast_fp16, y = var_2778_to_fp16)[name = string("op_2779_cast_fp16")];
            fp32 norm_epsilon_0 = const()[name = string("norm_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 128, 1]> norm_cast_fp16 = rsqrt(epsilon = norm_epsilon_0, x = var_2779_cast_fp16)[name = string("norm_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> var_2781_cast_fp16 = mul(x = x_357_cast_fp16, y = norm_cast_fp16)[name = string("op_2781_cast_fp16")];
            tensor<fp16, [1024]> layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210903296)))];
            tensor<fp16, [1, 128, 1024]> var_2782_cast_fp16 = mul(x = var_2781_cast_fp16, y = layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_13_mlp_gate_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210905408))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214051200))))[name = string("layers_13_mlp_gate_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_gate_proj_weight_to_fp16_palettized, x = var_2782_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> var_2792_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_2792_cast_fp16")];
            tensor<fp16, [3072, 1024]> layers_13_mlp_up_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214051776))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217197568))))[name = string("layers_13_mlp_up_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 3072]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_up_proj_weight_to_fp16_palettized, x = var_2782_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 128, 3072]> input_cast_fp16 = mul(x = var_2792_cast_fp16, y = linear_96_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [1024, 3072]> layers_13_mlp_down_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217198144))), lut = tensor<fp16, [1, 1, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220343936))))[name = string("layers_13_mlp_down_proj_weight_to_fp16_palettized")];
            tensor<fp16, [1, 128, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_mlp_down_proj_weight_to_fp16_palettized, x = input_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 128, 1024]> hidden_state = add(x = x_357_cast_fp16, y = linear_97_cast_fp16)[name = string("op_2798_cast_fp16")];
        } -> (hidden_state);
}