| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.22.1"}})] |
| { |
| func main<ios18>(tensor<fp16, [1, 1025, 1, 1]> causal_mask, tensor<fp16, [3, 1024, 3]> conv_state_in, tensor<int32, [1, 1]> input_ids, tensor<fp16, [2, 1, 512, 1, 1024]> kv_cache_in, tensor<int32, [1]> position_ids, tensor<fp16, [1, 1, 1024, 1]> update_mask) { |
| fp16 attn_scale = const()[name = string("attn_scale"), val = fp16(0x1p-3)]; |
| tensor<fp16, [64]> layers_2_self_attn_k_layernorm_weight = const()[name = string("layers_2_self_attn_k_layernorm_weight"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| tensor<fp16, [64]> layers_2_self_attn_q_layernorm_weight = const()[name = string("layers_2_self_attn_q_layernorm_weight"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256)))]; |
| tensor<fp16, [1024]> layers_0_operator_norm_weight = const()[name = string("layers_0_operator_norm_weight"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448)))]; |
| tensor<fp16, [2048, 64]> sin_cached_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2560))), lut = tensor<fp16, [64, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68160))))[name = string("sin_cached_palettized")]; |
| tensor<fp16, [2048, 64]> cos_cached_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [2048, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70272))), lut = tensor<fp16, [64, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135872))))[name = string("cos_cached_palettized")]; |
| tensor<fp16, [3072, 1024, 1, 1]> layers_0_conv_in_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137984))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1710912))))[name = string("layers_0_conv_in_proj_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_0_feed_forward_w1_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1714048))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4073408))))[name = string("layers_0_feed_forward_w1_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_0_feed_forward_w3_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4078080))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6437440))))[name = string("layers_0_feed_forward_w3_weight_palettized")]; |
| tensor<fp16, [1024, 4608, 1, 1]> layers_0_feed_forward_w2_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 4608, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6442112))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8801472))))[name = string("layers_0_feed_forward_w2_weight_palettized")]; |
| tensor<fp16, [3072, 1024, 1, 1]> layers_1_conv_in_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8802560))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10375488))))[name = string("layers_1_conv_in_proj_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_1_feed_forward_w1_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10378624))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12737984))))[name = string("layers_1_feed_forward_w1_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_1_feed_forward_w3_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12742656))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15102016))))[name = string("layers_1_feed_forward_w3_weight_palettized")]; |
| tensor<fp16, [1024, 4608, 1, 1]> layers_1_feed_forward_w2_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 4608, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15106688))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17466048))))[name = string("layers_1_feed_forward_w2_weight_palettized")]; |
| tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_q_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17467136))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17991488))))[name = string("layers_2_self_attn_q_proj_weight_palettized")]; |
| tensor<fp16, [512, 1024, 1, 1]> layers_2_self_attn_k_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17992576))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18254784))))[name = string("layers_2_self_attn_k_proj_weight_palettized")]; |
| tensor<fp16, [512, 1024, 1, 1]> layers_2_self_attn_v_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [512, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18255360))), lut = tensor<fp16, [16, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18517568))))[name = string("layers_2_self_attn_v_proj_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_2_feed_forward_w1_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18518144))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20877504))))[name = string("layers_2_feed_forward_w1_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_2_feed_forward_w3_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20882176))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23241536))))[name = string("layers_2_feed_forward_w3_weight_palettized")]; |
| tensor<fp16, [1024, 4608, 1, 1]> layers_2_feed_forward_w2_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 4608, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23246208))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25605568))))[name = string("layers_2_feed_forward_w2_weight_palettized")]; |
| tensor<fp16, [3072, 1024, 1, 1]> layers_3_conv_in_proj_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25606656))), lut = tensor<fp16, [96, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27179584))))[name = string("layers_3_conv_in_proj_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_3_feed_forward_w1_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27182720))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29542080))))[name = string("layers_3_feed_forward_w1_weight_palettized")]; |
| tensor<fp16, [4608, 1024, 1, 1]> layers_3_feed_forward_w3_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [4608, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29546752))), lut = tensor<fp16, [144, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31906112))))[name = string("layers_3_feed_forward_w3_weight_palettized")]; |
| tensor<fp16, [1024, 4608, 1, 1]> layers_3_feed_forward_w2_weight_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 4608, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31910784))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34270144))))[name = string("layers_3_feed_forward_w2_weight_palettized")]; |
| int32 var_156_batch_dims_0 = const()[name = string("op_156_batch_dims_0"), val = int32(0)]; |
| bool var_156_validate_indices_0 = const()[name = string("op_156_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [65536, 1024]> embed_tokens_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [65536, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34271232))), lut = tensor<fp16, [2048, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67825728))))[name = string("embed_tokens_weight_to_fp16_palettized")]; |
| int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)]; |
| tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")]; |
| int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(65536)]; |
| tensor<int32, [1, 1]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")]; |
| tensor<int32, [1, 1]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")]; |
| int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)]; |
| tensor<bool, [1, 1]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")]; |
| int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(65536)]; |
| tensor<int32, [1, 1]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")]; |
| tensor<int32, [1, 1]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")]; |
| int32 op_156_cast_fp16_axis_0 = const()[name = string("op_156_cast_fp16_axis_0"), val = int32(0)]; |
| tensor<fp16, [1, 1, 1024]> op_156_cast_fp16 = gather(axis = op_156_cast_fp16_axis_0, batch_dims = var_156_batch_dims_0, indices = select_0_1, validate_indices = var_156_validate_indices_0, x = embed_tokens_weight_to_fp16_palettized)[name = string("op_156_cast_fp16")]; |
| int32 var_162 = const()[name = string("op_162"), val = int32(0)]; |
| int32 var_163_batch_dims_0 = const()[name = string("op_163_batch_dims_0"), val = int32(0)]; |
| bool var_163_validate_indices_0 = const()[name = string("op_163_validate_indices_0"), val = bool(false)]; |
| string position_ids_to_uint16_dtype_0 = const()[name = string("position_ids_to_uint16_dtype_0"), val = string("uint16")]; |
| tensor<uint16, [1]> position_ids_to_uint16 = cast(dtype = position_ids_to_uint16_dtype_0, x = position_ids)[name = string("cast_0")]; |
| tensor<fp16, [1, 64]> var_163_cast_uint16 = gather(axis = var_162, batch_dims = var_163_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_163_validate_indices_0, x = cos_cached_palettized)[name = string("op_163_cast_uint16")]; |
| tensor<int32, [4]> var_168 = const()[name = string("op_168"), val = tensor<int32, [4]>([1, 1, 1, 64])]; |
| tensor<fp16, [1, 1, 1, 64]> cos = reshape(shape = var_168, x = var_163_cast_uint16)[name = string("cos")]; |
| int32 var_170 = const()[name = string("op_170"), val = int32(0)]; |
| int32 var_171_batch_dims_0 = const()[name = string("op_171_batch_dims_0"), val = int32(0)]; |
| bool var_171_validate_indices_0 = const()[name = string("op_171_validate_indices_0"), val = bool(false)]; |
| tensor<fp16, [1, 64]> var_171_cast_uint16 = gather(axis = var_170, batch_dims = var_171_batch_dims_0, indices = position_ids_to_uint16, validate_indices = var_171_validate_indices_0, x = sin_cached_palettized)[name = string("op_171_cast_uint16")]; |
| tensor<int32, [4]> var_176 = const()[name = string("op_176"), val = tensor<int32, [4]>([1, 1, 1, 64])]; |
| tensor<fp16, [1, 1, 1, 64]> sin = reshape(shape = var_176, x = var_171_cast_uint16)[name = string("sin")]; |
| fp16 const_0_promoted = const()[name = string("const_0_promoted"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_178 = mul(x = op_156_cast_fp16, y = const_0_promoted)[name = string("op_178")]; |
| int32 var_180 = const()[name = string("op_180"), val = int32(-1)]; |
| bool input_1_interleave_0 = const()[name = string("input_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_1 = concat(axis = var_180, interleave = input_1_interleave_0, values = (op_156_cast_fp16, var_178))[name = string("input_1")]; |
| tensor<int32, [1]> normed_1_axes_0 = const()[name = string("normed_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_186_to_fp16 = const()[name = string("op_186_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_1_cast_fp16 = layer_norm(axes = normed_1_axes_0, epsilon = var_186_to_fp16, x = input_1)[name = string("normed_1_cast_fp16")]; |
| tensor<int32, [2]> var_189_split_sizes_0 = const()[name = string("op_189_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_189_axis_0 = const()[name = string("op_189_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_189_0, tensor<fp16, [1, 1, 1024]> var_189_1 = split(axis = var_189_axis_0, split_sizes = var_189_split_sizes_0, x = normed_1_cast_fp16)[name = string("op_189")]; |
| tensor<fp16, [1, 1, 1024]> hidden_states_1 = mul(x = var_189_0, y = layers_0_operator_norm_weight)[name = string("hidden_states_1")]; |
| tensor<int32, [3]> var_195 = const()[name = string("op_195"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_198_axes_0 = const()[name = string("op_198_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_196 = transpose(perm = var_195, x = hidden_states_1)[name = string("transpose_20")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_198 = expand_dims(axes = var_198_axes_0, x = var_196)[name = string("op_198")]; |
| string BCx_1_pad_type_0 = const()[name = string("BCx_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> BCx_1_strides_0 = const()[name = string("BCx_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> BCx_1_pad_0 = const()[name = string("BCx_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> BCx_1_dilations_0 = const()[name = string("BCx_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 BCx_1_groups_0 = const()[name = string("BCx_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 3072, 1, 1]> BCx_1 = conv(dilations = BCx_1_dilations_0, groups = BCx_1_groups_0, pad = BCx_1_pad_0, pad_type = BCx_1_pad_type_0, strides = BCx_1_strides_0, weight = layers_0_conv_in_proj_weight_palettized, x = var_198)[name = string("BCx_1")]; |
| tensor<int32, [3]> var_215_split_sizes_0 = const()[name = string("op_215_split_sizes_0"), val = tensor<int32, [3]>([1024, 1024, 1024])]; |
| int32 var_215_axis_0 = const()[name = string("op_215_axis_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> var_215_0, tensor<fp16, [1, 1024, 1, 1]> var_215_1, tensor<fp16, [1, 1024, 1, 1]> var_215_2 = split(axis = var_215_axis_0, split_sizes = var_215_split_sizes_0, x = BCx_1)[name = string("op_215")]; |
| tensor<fp16, [1, 1024, 1, 1]> Bx_1 = mul(x = var_215_0, y = var_215_2)[name = string("Bx_1")]; |
| tensor<int32, [3]> var_221_begin_0 = const()[name = string("op_221_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> var_221_end_0 = const()[name = string("op_221_end_0"), val = tensor<int32, [3]>([1, 1024, 3])]; |
| tensor<bool, [3]> var_221_end_mask_0 = const()[name = string("op_221_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> var_221_squeeze_mask_0 = const()[name = string("op_221_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp16, [1024, 3]> var_221_cast_fp16 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, squeeze_mask = var_221_squeeze_mask_0, x = conv_state_in)[name = string("op_221_cast_fp16")]; |
| tensor<int32, [1]> var_223_axes_0 = const()[name = string("op_223_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1, 1024, 3]> var_223_cast_fp16 = expand_dims(axes = var_223_axes_0, x = var_221_cast_fp16)[name = string("op_223_cast_fp16")]; |
| tensor<int32, [1]> slot_1_axes_0 = const()[name = string("slot_1_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1, 3]> slot_1_cast_fp16 = expand_dims(axes = slot_1_axes_0, x = var_223_cast_fp16)[name = string("slot_1_cast_fp16")]; |
| tensor<int32, [4]> live_tail_1_begin_0 = const()[name = string("live_tail_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])]; |
| tensor<int32, [4]> live_tail_1_end_0 = const()[name = string("live_tail_1_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1])]; |
| tensor<bool, [4]> live_tail_1_end_mask_0 = const()[name = string("live_tail_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])]; |
| tensor<fp16, [1, 1024, 1, 2]> live_tail_1_cast_fp16 = slice_by_index(begin = live_tail_1_begin_0, end = live_tail_1_end_0, end_mask = live_tail_1_end_mask_0, x = slot_1_cast_fp16)[name = string("live_tail_1_cast_fp16")]; |
| int32 var_232 = const()[name = string("op_232"), val = int32(-1)]; |
| bool new_state_1_interleave_0 = const()[name = string("new_state_1_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1024, 1, 3]> new_state_1_cast_fp16 = concat(axis = var_232, interleave = new_state_1_interleave_0, values = (live_tail_1_cast_fp16, Bx_1))[name = string("new_state_1_cast_fp16")]; |
| tensor<int32, [1]> var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1024, 1, 3]> var_235_cast_fp16 = squeeze(axes = var_235_axes_0, x = new_state_1_cast_fp16)[name = string("op_235_cast_fp16")]; |
| tensor<int32, [1]> var_237_axes_0 = const()[name = string("op_237_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1024, 3]> var_237_cast_fp16 = squeeze(axes = var_237_axes_0, x = var_235_cast_fp16)[name = string("op_237_cast_fp16")]; |
| string conv_out_1_pad_type_0 = const()[name = string("conv_out_1_pad_type_0"), val = string("valid")]; |
| int32 conv_out_1_groups_0 = const()[name = string("conv_out_1_groups_0"), val = int32(1024)]; |
| tensor<int32, [2]> conv_out_1_strides_0 = const()[name = string("conv_out_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> conv_out_1_pad_0 = const()[name = string("conv_out_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> conv_out_1_dilations_0 = const()[name = string("conv_out_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<fp16, [1024, 1, 1, 3]> layers_0_conv_conv_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67891328))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67892928))))[name = string("layers_0_conv_conv_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> conv_out_1_cast_fp16 = conv(dilations = conv_out_1_dilations_0, groups = conv_out_1_groups_0, pad = conv_out_1_pad_0, pad_type = conv_out_1_pad_type_0, strides = conv_out_1_strides_0, weight = layers_0_conv_conv_weight_promoted_to_fp16_palettized, x = new_state_1_cast_fp16)[name = string("conv_out_1_cast_fp16")]; |
| tensor<fp16, [1, 1024, 1, 1]> input_5_cast_fp16 = mul(x = var_215_1, y = conv_out_1_cast_fp16)[name = string("input_5_cast_fp16")]; |
| string y_1_pad_type_0 = const()[name = string("y_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> y_1_strides_0 = const()[name = string("y_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> y_1_pad_0 = const()[name = string("y_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> y_1_dilations_0 = const()[name = string("y_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 y_1_groups_0 = const()[name = string("y_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 1024, 1, 1]> layers_0_conv_out_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67894016))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68418368))))[name = string("layers_0_conv_out_proj_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> y_1_cast_fp16 = conv(dilations = y_1_dilations_0, groups = y_1_groups_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = layers_0_conv_out_proj_weight_promoted_to_fp16_palettized, x = input_5_cast_fp16)[name = string("y_1_cast_fp16")]; |
| tensor<int32, [1]> var_263_axes_0 = const()[name = string("op_263_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_263_cast_fp16 = squeeze(axes = var_263_axes_0, x = y_1_cast_fp16)[name = string("op_263_cast_fp16")]; |
| tensor<int32, [3]> var_267 = const()[name = string("op_267"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> op_out_1_cast_fp16 = transpose(perm = var_267, x = var_263_cast_fp16)[name = string("transpose_19")]; |
| tensor<fp16, [1, 1, 1024]> x_3_cast_fp16 = add(x = op_156_cast_fp16, y = op_out_1_cast_fp16)[name = string("x_3_cast_fp16")]; |
| fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_271_cast_fp16 = mul(x = x_3_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_271_cast_fp16")]; |
| int32 var_273 = const()[name = string("op_273"), val = int32(-1)]; |
| bool input_7_interleave_0 = const()[name = string("input_7_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_7_cast_fp16 = concat(axis = var_273, interleave = input_7_interleave_0, values = (x_3_cast_fp16, var_271_cast_fp16))[name = string("input_7_cast_fp16")]; |
| tensor<int32, [1]> normed_3_axes_0 = const()[name = string("normed_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_279_to_fp16 = const()[name = string("op_279_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_3_cast_fp16 = layer_norm(axes = normed_3_axes_0, epsilon = var_279_to_fp16, x = input_7_cast_fp16)[name = string("normed_3_cast_fp16")]; |
| tensor<int32, [2]> var_282_split_sizes_0 = const()[name = string("op_282_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_282_axis_0 = const()[name = string("op_282_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_282_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_282_cast_fp16_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = normed_3_cast_fp16)[name = string("op_282_cast_fp16")]; |
| tensor<fp16, [1024]> layers_0_ffn_norm_weight_promoted_to_fp16 = const()[name = string("layers_0_ffn_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68419456)))]; |
| tensor<fp16, [1, 1, 1024]> normed_5_cast_fp16 = mul(x = var_282_cast_fp16_0, y = layers_0_ffn_norm_weight_promoted_to_fp16)[name = string("normed_5_cast_fp16")]; |
| tensor<int32, [3]> var_288 = const()[name = string("op_288"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_291_axes_0 = const()[name = string("op_291_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_289_cast_fp16 = transpose(perm = var_288, x = normed_5_cast_fp16)[name = string("transpose_18")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_291_cast_fp16 = expand_dims(axes = var_291_axes_0, x = var_289_cast_fp16)[name = string("op_291_cast_fp16")]; |
| string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> input_11 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_0_feed_forward_w1_weight_palettized, x = var_291_cast_fp16)[name = string("input_11")]; |
| string b_1_pad_type_0 = const()[name = string("b_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> b_1_strides_0 = const()[name = string("b_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> b_1_pad_0 = const()[name = string("b_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> b_1_dilations_0 = const()[name = string("b_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 b_1_groups_0 = const()[name = string("b_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> b_1 = conv(dilations = b_1_dilations_0, groups = b_1_groups_0, pad = b_1_pad_0, pad_type = b_1_pad_type_0, strides = b_1_strides_0, weight = layers_0_feed_forward_w3_weight_palettized, x = var_291_cast_fp16)[name = string("b_1")]; |
| tensor<fp16, [1, 4608, 1, 1]> var_319 = silu(x = input_11)[name = string("op_319")]; |
| tensor<fp16, [1, 4608, 1, 1]> input_13 = mul(x = var_319, y = b_1)[name = string("input_13")]; |
| string mlp_1_pad_type_0 = const()[name = string("mlp_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> mlp_1_strides_0 = const()[name = string("mlp_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> mlp_1_pad_0 = const()[name = string("mlp_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> mlp_1_dilations_0 = const()[name = string("mlp_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 mlp_1_groups_0 = const()[name = string("mlp_1_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> mlp_1 = conv(dilations = mlp_1_dilations_0, groups = mlp_1_groups_0, pad = mlp_1_pad_0, pad_type = mlp_1_pad_type_0, strides = mlp_1_strides_0, weight = layers_0_feed_forward_w2_weight_palettized, x = input_13)[name = string("mlp_1")]; |
| tensor<int32, [1]> var_333_axes_0 = const()[name = string("op_333_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_333 = squeeze(axes = var_333_axes_0, x = mlp_1)[name = string("op_333")]; |
| tensor<int32, [3]> var_337 = const()[name = string("op_337"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> mlp_3 = transpose(perm = var_337, x = var_333)[name = string("transpose_17")]; |
| tensor<fp16, [1, 1, 1024]> x_5_cast_fp16 = add(x = x_3_cast_fp16, y = mlp_3)[name = string("x_5_cast_fp16")]; |
| fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_341_cast_fp16 = mul(x = x_5_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_341_cast_fp16")]; |
| int32 var_343 = const()[name = string("op_343"), val = int32(-1)]; |
| bool input_15_interleave_0 = const()[name = string("input_15_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_15_cast_fp16 = concat(axis = var_343, interleave = input_15_interleave_0, values = (x_5_cast_fp16, var_341_cast_fp16))[name = string("input_15_cast_fp16")]; |
| tensor<int32, [1]> normed_7_axes_0 = const()[name = string("normed_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_349_to_fp16 = const()[name = string("op_349_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_7_cast_fp16 = layer_norm(axes = normed_7_axes_0, epsilon = var_349_to_fp16, x = input_15_cast_fp16)[name = string("normed_7_cast_fp16")]; |
| tensor<int32, [2]> var_352_split_sizes_0 = const()[name = string("op_352_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_352_axis_0 = const()[name = string("op_352_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_352_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_352_cast_fp16_1 = split(axis = var_352_axis_0, split_sizes = var_352_split_sizes_0, x = normed_7_cast_fp16)[name = string("op_352_cast_fp16")]; |
| tensor<fp16, [1024]> layers_1_operator_norm_weight_promoted_to_fp16 = const()[name = string("layers_1_operator_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68421568)))]; |
| tensor<fp16, [1, 1, 1024]> hidden_states_3_cast_fp16 = mul(x = var_352_cast_fp16_0, y = layers_1_operator_norm_weight_promoted_to_fp16)[name = string("hidden_states_3_cast_fp16")]; |
| tensor<int32, [3]> var_358 = const()[name = string("op_358"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_361_axes_0 = const()[name = string("op_361_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_359_cast_fp16 = transpose(perm = var_358, x = hidden_states_3_cast_fp16)[name = string("transpose_16")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_361_cast_fp16 = expand_dims(axes = var_361_axes_0, x = var_359_cast_fp16)[name = string("op_361_cast_fp16")]; |
| string BCx_3_pad_type_0 = const()[name = string("BCx_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> BCx_3_strides_0 = const()[name = string("BCx_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> BCx_3_pad_0 = const()[name = string("BCx_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> BCx_3_dilations_0 = const()[name = string("BCx_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 BCx_3_groups_0 = const()[name = string("BCx_3_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 3072, 1, 1]> BCx_3 = conv(dilations = BCx_3_dilations_0, groups = BCx_3_groups_0, pad = BCx_3_pad_0, pad_type = BCx_3_pad_type_0, strides = BCx_3_strides_0, weight = layers_1_conv_in_proj_weight_palettized, x = var_361_cast_fp16)[name = string("BCx_3")]; |
| tensor<int32, [3]> var_378_split_sizes_0 = const()[name = string("op_378_split_sizes_0"), val = tensor<int32, [3]>([1024, 1024, 1024])]; |
| int32 var_378_axis_0 = const()[name = string("op_378_axis_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> var_378_0, tensor<fp16, [1, 1024, 1, 1]> var_378_1, tensor<fp16, [1, 1024, 1, 1]> var_378_2 = split(axis = var_378_axis_0, split_sizes = var_378_split_sizes_0, x = BCx_3)[name = string("op_378")]; |
| tensor<fp16, [1, 1024, 1, 1]> Bx_3 = mul(x = var_378_0, y = var_378_2)[name = string("Bx_3")]; |
| tensor<int32, [3]> var_384_begin_0 = const()[name = string("op_384_begin_0"), val = tensor<int32, [3]>([1, 0, 0])]; |
| tensor<int32, [3]> var_384_end_0 = const()[name = string("op_384_end_0"), val = tensor<int32, [3]>([2, 1024, 3])]; |
| tensor<bool, [3]> var_384_end_mask_0 = const()[name = string("op_384_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> var_384_squeeze_mask_0 = const()[name = string("op_384_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp16, [1024, 3]> var_384_cast_fp16 = slice_by_index(begin = var_384_begin_0, end = var_384_end_0, end_mask = var_384_end_mask_0, squeeze_mask = var_384_squeeze_mask_0, x = conv_state_in)[name = string("op_384_cast_fp16")]; |
| tensor<int32, [1]> var_386_axes_0 = const()[name = string("op_386_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1, 1024, 3]> var_386_cast_fp16 = expand_dims(axes = var_386_axes_0, x = var_384_cast_fp16)[name = string("op_386_cast_fp16")]; |
| tensor<int32, [1]> slot_3_axes_0 = const()[name = string("slot_3_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1, 3]> slot_3_cast_fp16 = expand_dims(axes = slot_3_axes_0, x = var_386_cast_fp16)[name = string("slot_3_cast_fp16")]; |
| tensor<int32, [4]> live_tail_3_begin_0 = const()[name = string("live_tail_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])]; |
| tensor<int32, [4]> live_tail_3_end_0 = const()[name = string("live_tail_3_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1])]; |
| tensor<bool, [4]> live_tail_3_end_mask_0 = const()[name = string("live_tail_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])]; |
| tensor<fp16, [1, 1024, 1, 2]> live_tail_3_cast_fp16 = slice_by_index(begin = live_tail_3_begin_0, end = live_tail_3_end_0, end_mask = live_tail_3_end_mask_0, x = slot_3_cast_fp16)[name = string("live_tail_3_cast_fp16")]; |
| int32 var_395 = const()[name = string("op_395"), val = int32(-1)]; |
| bool new_state_3_interleave_0 = const()[name = string("new_state_3_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1024, 1, 3]> new_state_3_cast_fp16 = concat(axis = var_395, interleave = new_state_3_interleave_0, values = (live_tail_3_cast_fp16, Bx_3))[name = string("new_state_3_cast_fp16")]; |
| tensor<int32, [1]> var_398_axes_0 = const()[name = string("op_398_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1024, 1, 3]> var_398_cast_fp16 = squeeze(axes = var_398_axes_0, x = new_state_3_cast_fp16)[name = string("op_398_cast_fp16")]; |
| tensor<int32, [1]> var_400_axes_0 = const()[name = string("op_400_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1024, 3]> var_400_cast_fp16 = squeeze(axes = var_400_axes_0, x = var_398_cast_fp16)[name = string("op_400_cast_fp16")]; |
| string conv_out_3_pad_type_0 = const()[name = string("conv_out_3_pad_type_0"), val = string("valid")]; |
| int32 conv_out_3_groups_0 = const()[name = string("conv_out_3_groups_0"), val = int32(1024)]; |
| tensor<int32, [2]> conv_out_3_strides_0 = const()[name = string("conv_out_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> conv_out_3_pad_0 = const()[name = string("conv_out_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> conv_out_3_dilations_0 = const()[name = string("conv_out_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<fp16, [1024, 1, 1, 3]> layers_1_conv_conv_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68423680))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68425280))))[name = string("layers_1_conv_conv_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> conv_out_3_cast_fp16 = conv(dilations = conv_out_3_dilations_0, groups = conv_out_3_groups_0, pad = conv_out_3_pad_0, pad_type = conv_out_3_pad_type_0, strides = conv_out_3_strides_0, weight = layers_1_conv_conv_weight_promoted_to_fp16_palettized, x = new_state_3_cast_fp16)[name = string("conv_out_3_cast_fp16")]; |
| tensor<fp16, [1, 1024, 1, 1]> input_19_cast_fp16 = mul(x = var_378_1, y = conv_out_3_cast_fp16)[name = string("input_19_cast_fp16")]; |
| string y_3_pad_type_0 = const()[name = string("y_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> y_3_strides_0 = const()[name = string("y_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> y_3_pad_0 = const()[name = string("y_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> y_3_dilations_0 = const()[name = string("y_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 y_3_groups_0 = const()[name = string("y_3_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 1024, 1, 1]> layers_1_conv_out_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68426368))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68950720))))[name = string("layers_1_conv_out_proj_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> y_3_cast_fp16 = conv(dilations = y_3_dilations_0, groups = y_3_groups_0, pad = y_3_pad_0, pad_type = y_3_pad_type_0, strides = y_3_strides_0, weight = layers_1_conv_out_proj_weight_promoted_to_fp16_palettized, x = input_19_cast_fp16)[name = string("y_3_cast_fp16")]; |
| tensor<int32, [1]> var_426_axes_0 = const()[name = string("op_426_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_426_cast_fp16 = squeeze(axes = var_426_axes_0, x = y_3_cast_fp16)[name = string("op_426_cast_fp16")]; |
| tensor<int32, [3]> var_430 = const()[name = string("op_430"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> op_out_3_cast_fp16 = transpose(perm = var_430, x = var_426_cast_fp16)[name = string("transpose_15")]; |
| tensor<fp16, [1, 1, 1024]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = op_out_3_cast_fp16)[name = string("x_7_cast_fp16")]; |
| fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_434_cast_fp16 = mul(x = x_7_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_434_cast_fp16")]; |
| int32 var_436 = const()[name = string("op_436"), val = int32(-1)]; |
| bool input_21_interleave_0 = const()[name = string("input_21_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_21_cast_fp16 = concat(axis = var_436, interleave = input_21_interleave_0, values = (x_7_cast_fp16, var_434_cast_fp16))[name = string("input_21_cast_fp16")]; |
| tensor<int32, [1]> normed_9_axes_0 = const()[name = string("normed_9_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_442_to_fp16 = const()[name = string("op_442_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_9_cast_fp16 = layer_norm(axes = normed_9_axes_0, epsilon = var_442_to_fp16, x = input_21_cast_fp16)[name = string("normed_9_cast_fp16")]; |
| tensor<int32, [2]> var_445_split_sizes_0 = const()[name = string("op_445_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_445_axis_0 = const()[name = string("op_445_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_445_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_445_cast_fp16_1 = split(axis = var_445_axis_0, split_sizes = var_445_split_sizes_0, x = normed_9_cast_fp16)[name = string("op_445_cast_fp16")]; |
| tensor<fp16, [1024]> layers_1_ffn_norm_weight_promoted_to_fp16 = const()[name = string("layers_1_ffn_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68951808)))]; |
| tensor<fp16, [1, 1, 1024]> normed_11_cast_fp16 = mul(x = var_445_cast_fp16_0, y = layers_1_ffn_norm_weight_promoted_to_fp16)[name = string("normed_11_cast_fp16")]; |
| tensor<int32, [3]> var_451 = const()[name = string("op_451"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_454_axes_0 = const()[name = string("op_454_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_452_cast_fp16 = transpose(perm = var_451, x = normed_11_cast_fp16)[name = string("transpose_14")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_454_cast_fp16 = expand_dims(axes = var_454_axes_0, x = var_452_cast_fp16)[name = string("op_454_cast_fp16")]; |
| string input_25_pad_type_0 = const()[name = string("input_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_25_strides_0 = const()[name = string("input_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_25_pad_0 = const()[name = string("input_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_25_dilations_0 = const()[name = string("input_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_25_groups_0 = const()[name = string("input_25_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> input_25 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_1_feed_forward_w1_weight_palettized, x = var_454_cast_fp16)[name = string("input_25")]; |
| string b_3_pad_type_0 = const()[name = string("b_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> b_3_strides_0 = const()[name = string("b_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> b_3_pad_0 = const()[name = string("b_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> b_3_dilations_0 = const()[name = string("b_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 b_3_groups_0 = const()[name = string("b_3_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> b_3 = conv(dilations = b_3_dilations_0, groups = b_3_groups_0, pad = b_3_pad_0, pad_type = b_3_pad_type_0, strides = b_3_strides_0, weight = layers_1_feed_forward_w3_weight_palettized, x = var_454_cast_fp16)[name = string("b_3")]; |
| tensor<fp16, [1, 4608, 1, 1]> var_482 = silu(x = input_25)[name = string("op_482")]; |
| tensor<fp16, [1, 4608, 1, 1]> input_27 = mul(x = var_482, y = b_3)[name = string("input_27")]; |
| string mlp_5_pad_type_0 = const()[name = string("mlp_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> mlp_5_strides_0 = const()[name = string("mlp_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> mlp_5_pad_0 = const()[name = string("mlp_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> mlp_5_dilations_0 = const()[name = string("mlp_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 mlp_5_groups_0 = const()[name = string("mlp_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> mlp_5 = conv(dilations = mlp_5_dilations_0, groups = mlp_5_groups_0, pad = mlp_5_pad_0, pad_type = mlp_5_pad_type_0, strides = mlp_5_strides_0, weight = layers_1_feed_forward_w2_weight_palettized, x = input_27)[name = string("mlp_5")]; |
| tensor<int32, [1]> var_496_axes_0 = const()[name = string("op_496_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_496 = squeeze(axes = var_496_axes_0, x = mlp_5)[name = string("op_496")]; |
| tensor<int32, [3]> var_500 = const()[name = string("op_500"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> mlp_7 = transpose(perm = var_500, x = var_496)[name = string("transpose_13")]; |
| tensor<fp16, [1, 1, 1024]> x_9_cast_fp16 = add(x = x_7_cast_fp16, y = mlp_7)[name = string("x_9_cast_fp16")]; |
| fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_504_cast_fp16 = mul(x = x_9_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_504_cast_fp16")]; |
| int32 var_506 = const()[name = string("op_506"), val = int32(-1)]; |
| bool input_29_interleave_0 = const()[name = string("input_29_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_29_cast_fp16 = concat(axis = var_506, interleave = input_29_interleave_0, values = (x_9_cast_fp16, var_504_cast_fp16))[name = string("input_29_cast_fp16")]; |
| tensor<int32, [1]> normed_13_axes_0 = const()[name = string("normed_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_512_to_fp16 = const()[name = string("op_512_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_13_cast_fp16 = layer_norm(axes = normed_13_axes_0, epsilon = var_512_to_fp16, x = input_29_cast_fp16)[name = string("normed_13_cast_fp16")]; |
| tensor<int32, [2]> var_515_split_sizes_0 = const()[name = string("op_515_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_515_axis_0 = const()[name = string("op_515_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_515_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_515_cast_fp16_1 = split(axis = var_515_axis_0, split_sizes = var_515_split_sizes_0, x = normed_13_cast_fp16)[name = string("op_515_cast_fp16")]; |
| tensor<fp16, [1024]> layers_2_operator_norm_weight_promoted_to_fp16 = const()[name = string("layers_2_operator_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68953920)))]; |
| tensor<fp16, [1, 1, 1024]> hidden_states_5_cast_fp16 = mul(x = var_515_cast_fp16_0, y = layers_2_operator_norm_weight_promoted_to_fp16)[name = string("hidden_states_5_cast_fp16")]; |
| tensor<int32, [3]> var_521 = const()[name = string("op_521"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_524_axes_0 = const()[name = string("op_524_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_522_cast_fp16 = transpose(perm = var_521, x = hidden_states_5_cast_fp16)[name = string("transpose_12")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_524_cast_fp16 = expand_dims(axes = var_524_axes_0, x = var_522_cast_fp16)[name = string("op_524_cast_fp16")]; |
| string var_540_pad_type_0 = const()[name = string("op_540_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_540_strides_0 = const()[name = string("op_540_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_540_pad_0 = const()[name = string("op_540_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_540_dilations_0 = const()[name = string("op_540_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_540_groups_0 = const()[name = string("op_540_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> var_540 = conv(dilations = var_540_dilations_0, groups = var_540_groups_0, pad = var_540_pad_0, pad_type = var_540_pad_type_0, strides = var_540_strides_0, weight = layers_2_self_attn_q_proj_weight_palettized, x = var_524_cast_fp16)[name = string("op_540")]; |
| tensor<int32, [4]> var_545 = const()[name = string("op_545"), val = tensor<int32, [4]>([1, 16, 64, 1])]; |
| tensor<fp16, [1, 16, 64, 1]> var_546 = reshape(shape = var_545, x = var_540)[name = string("op_546")]; |
| tensor<int32, [4]> var_551 = const()[name = string("op_551"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| string var_568_pad_type_0 = const()[name = string("op_568_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_568_strides_0 = const()[name = string("op_568_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_568_pad_0 = const()[name = string("op_568_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_568_dilations_0 = const()[name = string("op_568_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_568_groups_0 = const()[name = string("op_568_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 512, 1, 1]> var_568 = conv(dilations = var_568_dilations_0, groups = var_568_groups_0, pad = var_568_pad_0, pad_type = var_568_pad_type_0, strides = var_568_strides_0, weight = layers_2_self_attn_k_proj_weight_palettized, x = var_524_cast_fp16)[name = string("op_568")]; |
| tensor<int32, [4]> var_573 = const()[name = string("op_573"), val = tensor<int32, [4]>([1, 8, 64, 1])]; |
| tensor<fp16, [1, 8, 64, 1]> var_574 = reshape(shape = var_573, x = var_568)[name = string("op_574")]; |
| tensor<int32, [4]> var_579 = const()[name = string("op_579"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| string var_596_pad_type_0 = const()[name = string("op_596_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> var_596_strides_0 = const()[name = string("op_596_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> var_596_pad_0 = const()[name = string("op_596_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> var_596_dilations_0 = const()[name = string("op_596_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_596_groups_0 = const()[name = string("op_596_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 512, 1, 1]> var_596 = conv(dilations = var_596_dilations_0, groups = var_596_groups_0, pad = var_596_pad_0, pad_type = var_596_pad_type_0, strides = var_596_strides_0, weight = layers_2_self_attn_v_proj_weight_palettized, x = var_524_cast_fp16)[name = string("op_596")]; |
| fp16 const_5_promoted = const()[name = string("const_5_promoted"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 16, 1, 64]> var_552 = transpose(perm = var_551, x = var_546)[name = string("transpose_11")]; |
| tensor<fp16, [1, 16, 1, 64]> var_614 = mul(x = var_552, y = const_5_promoted)[name = string("op_614")]; |
| int32 var_616 = const()[name = string("op_616"), val = int32(-1)]; |
| bool input_33_interleave_0 = const()[name = string("input_33_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 1, 128]> input_33 = concat(axis = var_616, interleave = input_33_interleave_0, values = (var_552, var_614))[name = string("input_33")]; |
| tensor<int32, [1]> normed_15_axes_0 = const()[name = string("normed_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_622_to_fp16 = const()[name = string("op_622_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 16, 1, 128]> normed_15_cast_fp16 = layer_norm(axes = normed_15_axes_0, epsilon = var_622_to_fp16, x = input_33)[name = string("normed_15_cast_fp16")]; |
| tensor<int32, [2]> var_625_split_sizes_0 = const()[name = string("op_625_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_625_axis_0 = const()[name = string("op_625_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 16, 1, 64]> var_625_0, tensor<fp16, [1, 16, 1, 64]> var_625_1 = split(axis = var_625_axis_0, split_sizes = var_625_split_sizes_0, x = normed_15_cast_fp16)[name = string("op_625")]; |
| tensor<fp16, [1, 16, 1, 64]> q_1 = mul(x = var_625_0, y = layers_2_self_attn_q_layernorm_weight)[name = string("q_1")]; |
| fp16 const_6_promoted = const()[name = string("const_6_promoted"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 8, 1, 64]> var_580 = transpose(perm = var_579, x = var_574)[name = string("transpose_10")]; |
| tensor<fp16, [1, 8, 1, 64]> var_628 = mul(x = var_580, y = const_6_promoted)[name = string("op_628")]; |
| int32 var_630 = const()[name = string("op_630"), val = int32(-1)]; |
| bool input_35_interleave_0 = const()[name = string("input_35_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 8, 1, 128]> input_35 = concat(axis = var_630, interleave = input_35_interleave_0, values = (var_580, var_628))[name = string("input_35")]; |
| tensor<int32, [1]> normed_17_axes_0 = const()[name = string("normed_17_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_636_to_fp16 = const()[name = string("op_636_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 8, 1, 128]> normed_17_cast_fp16 = layer_norm(axes = normed_17_axes_0, epsilon = var_636_to_fp16, x = input_35)[name = string("normed_17_cast_fp16")]; |
| tensor<int32, [2]> var_639_split_sizes_0 = const()[name = string("op_639_split_sizes_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 var_639_axis_0 = const()[name = string("op_639_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 8, 1, 64]> var_639_0, tensor<fp16, [1, 8, 1, 64]> var_639_1 = split(axis = var_639_axis_0, split_sizes = var_639_split_sizes_0, x = normed_17_cast_fp16)[name = string("op_639")]; |
| tensor<fp16, [1, 8, 1, 64]> k_1 = mul(x = var_639_0, y = layers_2_self_attn_k_layernorm_weight)[name = string("k_1")]; |
| tensor<fp16, [1, 16, 1, 64]> var_642 = mul(x = q_1, y = cos)[name = string("op_642")]; |
| tensor<int32, [2]> var_643_split_sizes_0 = const()[name = string("op_643_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; |
| int32 var_643_axis_0 = const()[name = string("op_643_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 16, 1, 32]> var_643_0, tensor<fp16, [1, 16, 1, 32]> var_643_1 = split(axis = var_643_axis_0, split_sizes = var_643_split_sizes_0, x = q_1)[name = string("op_643")]; |
| fp16 const_7_promoted = const()[name = string("const_7_promoted"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 16, 1, 32]> var_645 = mul(x = var_643_1, y = const_7_promoted)[name = string("op_645")]; |
| int32 var_647 = const()[name = string("op_647"), val = int32(-1)]; |
| bool var_648_interleave_0 = const()[name = string("op_648_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 1, 64]> var_648 = concat(axis = var_647, interleave = var_648_interleave_0, values = (var_645, var_643_0))[name = string("op_648")]; |
| tensor<fp16, [1, 16, 1, 64]> var_649 = mul(x = var_648, y = sin)[name = string("op_649")]; |
| tensor<fp16, [1, 16, 1, 64]> q = add(x = var_642, y = var_649)[name = string("q")]; |
| tensor<fp16, [1, 8, 1, 64]> var_652 = mul(x = k_1, y = cos)[name = string("op_652")]; |
| tensor<int32, [2]> var_653_split_sizes_0 = const()[name = string("op_653_split_sizes_0"), val = tensor<int32, [2]>([32, 32])]; |
| int32 var_653_axis_0 = const()[name = string("op_653_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 8, 1, 32]> var_653_0, tensor<fp16, [1, 8, 1, 32]> var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = k_1)[name = string("op_653")]; |
| fp16 const_8_promoted = const()[name = string("const_8_promoted"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 8, 1, 32]> var_655 = mul(x = var_653_1, y = const_8_promoted)[name = string("op_655")]; |
| int32 var_657 = const()[name = string("op_657"), val = int32(-1)]; |
| bool var_658_interleave_0 = const()[name = string("op_658_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 8, 1, 64]> var_658 = concat(axis = var_657, interleave = var_658_interleave_0, values = (var_655, var_653_0))[name = string("op_658")]; |
| tensor<fp16, [1, 8, 1, 64]> var_659 = mul(x = var_658, y = sin)[name = string("op_659")]; |
| tensor<fp16, [1, 8, 1, 64]> k = add(x = var_652, y = var_659)[name = string("k")]; |
| tensor<int32, [5]> K_cache_begin_0 = const()[name = string("K_cache_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> K_cache_end_0 = const()[name = string("K_cache_end_0"), val = tensor<int32, [5]>([1, 1, 512, 1, 1024])]; |
| tensor<bool, [5]> K_cache_end_mask_0 = const()[name = string("K_cache_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> K_cache_squeeze_mask_0 = const()[name = string("K_cache_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [1, 512, 1, 1024]> K_cache_cast_fp16 = slice_by_index(begin = K_cache_begin_0, end = K_cache_end_0, end_mask = K_cache_end_mask_0, squeeze_mask = K_cache_squeeze_mask_0, x = kv_cache_in)[name = string("K_cache_cast_fp16")]; |
| tensor<int32, [5]> V_cache_begin_0 = const()[name = string("V_cache_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> V_cache_end_0 = const()[name = string("V_cache_end_0"), val = tensor<int32, [5]>([2, 1, 512, 1, 1024])]; |
| tensor<bool, [5]> V_cache_end_mask_0 = const()[name = string("V_cache_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> V_cache_squeeze_mask_0 = const()[name = string("V_cache_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp16, [1, 512, 1, 1024]> V_cache_cast_fp16 = slice_by_index(begin = V_cache_begin_0, end = V_cache_end_0, end_mask = V_cache_end_mask_0, squeeze_mask = V_cache_squeeze_mask_0, x = kv_cache_in)[name = string("V_cache_cast_fp16")]; |
| tensor<int32, [4]> var_672 = const()[name = string("op_672"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [4]> var_678 = const()[name = string("op_678"), val = tensor<int32, [4]>([1, 1024, 1, 1])]; |
| tensor<fp16, [1, 16, 64, 1]> var_673 = transpose(perm = var_672, x = q)[name = string("transpose_9")]; |
| tensor<fp16, [1, 1024, 1, 1]> query = reshape(shape = var_678, x = var_673)[name = string("query")]; |
| tensor<int32, [4]> var_684 = const()[name = string("op_684"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [4]> var_690 = const()[name = string("op_690"), val = tensor<int32, [4]>([1, 512, 1, 1])]; |
| tensor<fp16, [1, 8, 64, 1]> var_685 = transpose(perm = var_684, x = k)[name = string("transpose_8")]; |
| tensor<fp16, [1, 512, 1, 1]> k_slice = reshape(shape = var_690, x = var_685)[name = string("k_slice")]; |
| int32 var_705 = const()[name = string("op_705"), val = int32(-1)]; |
| bool key_interleave_0 = const()[name = string("key_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 512, 1, 1025]> key_cast_fp16 = concat(axis = var_705, interleave = key_interleave_0, values = (K_cache_cast_fp16, k_slice))[name = string("key_cast_fp16")]; |
| int32 var_708 = const()[name = string("op_708"), val = int32(-1)]; |
| bool var_709_interleave_0 = const()[name = string("op_709_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 512, 1, 1025]> var_709_cast_fp16 = concat(axis = var_708, interleave = var_709_interleave_0, values = (V_cache_cast_fp16, var_596))[name = string("op_709_cast_fp16")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_710 = mul(x = query, y = attn_scale)[name = string("op_710")]; |
| tensor<int32, [16]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68956032)))]; |
| int32 var_713_axis_0 = const()[name = string("op_713_axis_0"), val = int32(1)]; |
| tensor<fp16, [1, 64, 1, 1]> var_713_0, tensor<fp16, [1, 64, 1, 1]> var_713_1, tensor<fp16, [1, 64, 1, 1]> var_713_2, tensor<fp16, [1, 64, 1, 1]> var_713_3, tensor<fp16, [1, 64, 1, 1]> var_713_4, tensor<fp16, [1, 64, 1, 1]> var_713_5, tensor<fp16, [1, 64, 1, 1]> var_713_6, tensor<fp16, [1, 64, 1, 1]> var_713_7, tensor<fp16, [1, 64, 1, 1]> var_713_8, tensor<fp16, [1, 64, 1, 1]> var_713_9, tensor<fp16, [1, 64, 1, 1]> var_713_10, tensor<fp16, [1, 64, 1, 1]> var_713_11, tensor<fp16, [1, 64, 1, 1]> var_713_12, tensor<fp16, [1, 64, 1, 1]> var_713_13, tensor<fp16, [1, 64, 1, 1]> var_713_14, tensor<fp16, [1, 64, 1, 1]> var_713_15 = split(axis = var_713_axis_0, split_sizes = tile_0, x = var_710)[name = string("op_713")]; |
| tensor<int32, [4]> var_732_perm_0 = const()[name = string("op_732_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])]; |
| tensor<int32, [8]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])]; |
| int32 var_735_axis_0 = const()[name = string("op_735_axis_0"), val = int32(3)]; |
| tensor<fp16, [1, 1025, 1, 512]> var_732_cast_fp16 = transpose(perm = var_732_perm_0, x = key_cast_fp16)[name = string("transpose_7")]; |
| tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_0, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_1, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_2, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_3, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_4, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_5, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_6, tensor<fp16, [1, 1025, 1, 64]> var_735_cast_fp16_7 = split(axis = var_735_axis_0, split_sizes = tile_1, x = var_732_cast_fp16)[name = string("op_735_cast_fp16")]; |
| tensor<int32, [8]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])]; |
| int32 var_746_axis_0 = const()[name = string("op_746_axis_0"), val = int32(1)]; |
| tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_0, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_1, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_2, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_3, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_4, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_5, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_6, tensor<fp16, [1, 64, 1, 1025]> var_746_cast_fp16_7 = split(axis = var_746_axis_0, split_sizes = tile_2, x = var_709_cast_fp16)[name = string("op_746_cast_fp16")]; |
| string scores_1_equation_0 = const()[name = string("scores_1_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_1_cast_fp16 = einsum(equation = scores_1_equation_0, values = (var_735_cast_fp16_0, var_713_0))[name = string("scores_1_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_760_cast_fp16 = add(x = scores_1_cast_fp16, y = causal_mask)[name = string("op_760_cast_fp16")]; |
| int32 var_761 = const()[name = string("op_761"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_763_cast_fp16 = softmax(axis = var_761, x = var_760_cast_fp16)[name = string("op_763_cast_fp16")]; |
| string var_767_equation_0 = const()[name = string("op_767_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_767_cast_fp16 = einsum(equation = var_767_equation_0, values = (var_746_cast_fp16_0, var_763_cast_fp16))[name = string("op_767_cast_fp16")]; |
| string scores_3_equation_0 = const()[name = string("scores_3_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_3_cast_fp16 = einsum(equation = scores_3_equation_0, values = (var_735_cast_fp16_0, var_713_1))[name = string("scores_3_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_773_cast_fp16 = add(x = scores_3_cast_fp16, y = causal_mask)[name = string("op_773_cast_fp16")]; |
| int32 var_774 = const()[name = string("op_774"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_776_cast_fp16 = softmax(axis = var_774, x = var_773_cast_fp16)[name = string("op_776_cast_fp16")]; |
| string var_780_equation_0 = const()[name = string("op_780_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_780_cast_fp16 = einsum(equation = var_780_equation_0, values = (var_746_cast_fp16_0, var_776_cast_fp16))[name = string("op_780_cast_fp16")]; |
| string scores_5_equation_0 = const()[name = string("scores_5_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_5_cast_fp16 = einsum(equation = scores_5_equation_0, values = (var_735_cast_fp16_1, var_713_2))[name = string("scores_5_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_786_cast_fp16 = add(x = scores_5_cast_fp16, y = causal_mask)[name = string("op_786_cast_fp16")]; |
| int32 var_787 = const()[name = string("op_787"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_789_cast_fp16 = softmax(axis = var_787, x = var_786_cast_fp16)[name = string("op_789_cast_fp16")]; |
| string var_793_equation_0 = const()[name = string("op_793_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_793_cast_fp16 = einsum(equation = var_793_equation_0, values = (var_746_cast_fp16_1, var_789_cast_fp16))[name = string("op_793_cast_fp16")]; |
| string scores_7_equation_0 = const()[name = string("scores_7_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_7_cast_fp16 = einsum(equation = scores_7_equation_0, values = (var_735_cast_fp16_1, var_713_3))[name = string("scores_7_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_799_cast_fp16 = add(x = scores_7_cast_fp16, y = causal_mask)[name = string("op_799_cast_fp16")]; |
| int32 var_800 = const()[name = string("op_800"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_802_cast_fp16 = softmax(axis = var_800, x = var_799_cast_fp16)[name = string("op_802_cast_fp16")]; |
| string var_806_equation_0 = const()[name = string("op_806_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_806_cast_fp16 = einsum(equation = var_806_equation_0, values = (var_746_cast_fp16_1, var_802_cast_fp16))[name = string("op_806_cast_fp16")]; |
| string scores_9_equation_0 = const()[name = string("scores_9_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_9_cast_fp16 = einsum(equation = scores_9_equation_0, values = (var_735_cast_fp16_2, var_713_4))[name = string("scores_9_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_812_cast_fp16 = add(x = scores_9_cast_fp16, y = causal_mask)[name = string("op_812_cast_fp16")]; |
| int32 var_813 = const()[name = string("op_813"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_815_cast_fp16 = softmax(axis = var_813, x = var_812_cast_fp16)[name = string("op_815_cast_fp16")]; |
| string var_819_equation_0 = const()[name = string("op_819_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_819_cast_fp16 = einsum(equation = var_819_equation_0, values = (var_746_cast_fp16_2, var_815_cast_fp16))[name = string("op_819_cast_fp16")]; |
| string scores_11_equation_0 = const()[name = string("scores_11_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_11_cast_fp16 = einsum(equation = scores_11_equation_0, values = (var_735_cast_fp16_2, var_713_5))[name = string("scores_11_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_825_cast_fp16 = add(x = scores_11_cast_fp16, y = causal_mask)[name = string("op_825_cast_fp16")]; |
| int32 var_826 = const()[name = string("op_826"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_828_cast_fp16 = softmax(axis = var_826, x = var_825_cast_fp16)[name = string("op_828_cast_fp16")]; |
| string var_832_equation_0 = const()[name = string("op_832_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_832_cast_fp16 = einsum(equation = var_832_equation_0, values = (var_746_cast_fp16_2, var_828_cast_fp16))[name = string("op_832_cast_fp16")]; |
| string scores_13_equation_0 = const()[name = string("scores_13_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_13_cast_fp16 = einsum(equation = scores_13_equation_0, values = (var_735_cast_fp16_3, var_713_6))[name = string("scores_13_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_838_cast_fp16 = add(x = scores_13_cast_fp16, y = causal_mask)[name = string("op_838_cast_fp16")]; |
| int32 var_839 = const()[name = string("op_839"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_841_cast_fp16 = softmax(axis = var_839, x = var_838_cast_fp16)[name = string("op_841_cast_fp16")]; |
| string var_845_equation_0 = const()[name = string("op_845_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_845_cast_fp16 = einsum(equation = var_845_equation_0, values = (var_746_cast_fp16_3, var_841_cast_fp16))[name = string("op_845_cast_fp16")]; |
| string scores_15_equation_0 = const()[name = string("scores_15_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_15_cast_fp16 = einsum(equation = scores_15_equation_0, values = (var_735_cast_fp16_3, var_713_7))[name = string("scores_15_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_851_cast_fp16 = add(x = scores_15_cast_fp16, y = causal_mask)[name = string("op_851_cast_fp16")]; |
| int32 var_852 = const()[name = string("op_852"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_854_cast_fp16 = softmax(axis = var_852, x = var_851_cast_fp16)[name = string("op_854_cast_fp16")]; |
| string var_858_equation_0 = const()[name = string("op_858_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_858_cast_fp16 = einsum(equation = var_858_equation_0, values = (var_746_cast_fp16_3, var_854_cast_fp16))[name = string("op_858_cast_fp16")]; |
| string scores_17_equation_0 = const()[name = string("scores_17_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_17_cast_fp16 = einsum(equation = scores_17_equation_0, values = (var_735_cast_fp16_4, var_713_8))[name = string("scores_17_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_864_cast_fp16 = add(x = scores_17_cast_fp16, y = causal_mask)[name = string("op_864_cast_fp16")]; |
| int32 var_865 = const()[name = string("op_865"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_867_cast_fp16 = softmax(axis = var_865, x = var_864_cast_fp16)[name = string("op_867_cast_fp16")]; |
| string var_871_equation_0 = const()[name = string("op_871_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_871_cast_fp16 = einsum(equation = var_871_equation_0, values = (var_746_cast_fp16_4, var_867_cast_fp16))[name = string("op_871_cast_fp16")]; |
| string scores_19_equation_0 = const()[name = string("scores_19_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_19_cast_fp16 = einsum(equation = scores_19_equation_0, values = (var_735_cast_fp16_4, var_713_9))[name = string("scores_19_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_877_cast_fp16 = add(x = scores_19_cast_fp16, y = causal_mask)[name = string("op_877_cast_fp16")]; |
| int32 var_878 = const()[name = string("op_878"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_880_cast_fp16 = softmax(axis = var_878, x = var_877_cast_fp16)[name = string("op_880_cast_fp16")]; |
| string var_884_equation_0 = const()[name = string("op_884_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_884_cast_fp16 = einsum(equation = var_884_equation_0, values = (var_746_cast_fp16_4, var_880_cast_fp16))[name = string("op_884_cast_fp16")]; |
| string scores_21_equation_0 = const()[name = string("scores_21_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_21_cast_fp16 = einsum(equation = scores_21_equation_0, values = (var_735_cast_fp16_5, var_713_10))[name = string("scores_21_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_890_cast_fp16 = add(x = scores_21_cast_fp16, y = causal_mask)[name = string("op_890_cast_fp16")]; |
| int32 var_891 = const()[name = string("op_891"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_893_cast_fp16 = softmax(axis = var_891, x = var_890_cast_fp16)[name = string("op_893_cast_fp16")]; |
| string var_897_equation_0 = const()[name = string("op_897_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_897_cast_fp16 = einsum(equation = var_897_equation_0, values = (var_746_cast_fp16_5, var_893_cast_fp16))[name = string("op_897_cast_fp16")]; |
| string scores_23_equation_0 = const()[name = string("scores_23_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_23_cast_fp16 = einsum(equation = scores_23_equation_0, values = (var_735_cast_fp16_5, var_713_11))[name = string("scores_23_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_903_cast_fp16 = add(x = scores_23_cast_fp16, y = causal_mask)[name = string("op_903_cast_fp16")]; |
| int32 var_904 = const()[name = string("op_904"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_906_cast_fp16 = softmax(axis = var_904, x = var_903_cast_fp16)[name = string("op_906_cast_fp16")]; |
| string var_910_equation_0 = const()[name = string("op_910_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_910_cast_fp16 = einsum(equation = var_910_equation_0, values = (var_746_cast_fp16_5, var_906_cast_fp16))[name = string("op_910_cast_fp16")]; |
| string scores_25_equation_0 = const()[name = string("scores_25_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_25_cast_fp16 = einsum(equation = scores_25_equation_0, values = (var_735_cast_fp16_6, var_713_12))[name = string("scores_25_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_916_cast_fp16 = add(x = scores_25_cast_fp16, y = causal_mask)[name = string("op_916_cast_fp16")]; |
| int32 var_917 = const()[name = string("op_917"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_919_cast_fp16 = softmax(axis = var_917, x = var_916_cast_fp16)[name = string("op_919_cast_fp16")]; |
| string var_923_equation_0 = const()[name = string("op_923_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_923_cast_fp16 = einsum(equation = var_923_equation_0, values = (var_746_cast_fp16_6, var_919_cast_fp16))[name = string("op_923_cast_fp16")]; |
| string scores_27_equation_0 = const()[name = string("scores_27_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_27_cast_fp16 = einsum(equation = scores_27_equation_0, values = (var_735_cast_fp16_6, var_713_13))[name = string("scores_27_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_929_cast_fp16 = add(x = scores_27_cast_fp16, y = causal_mask)[name = string("op_929_cast_fp16")]; |
| int32 var_930 = const()[name = string("op_930"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_932_cast_fp16 = softmax(axis = var_930, x = var_929_cast_fp16)[name = string("op_932_cast_fp16")]; |
| string var_936_equation_0 = const()[name = string("op_936_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_936_cast_fp16 = einsum(equation = var_936_equation_0, values = (var_746_cast_fp16_6, var_932_cast_fp16))[name = string("op_936_cast_fp16")]; |
| string scores_29_equation_0 = const()[name = string("scores_29_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_29_cast_fp16 = einsum(equation = scores_29_equation_0, values = (var_735_cast_fp16_7, var_713_14))[name = string("scores_29_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_942_cast_fp16 = add(x = scores_29_cast_fp16, y = causal_mask)[name = string("op_942_cast_fp16")]; |
| int32 var_943 = const()[name = string("op_943"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_945_cast_fp16 = softmax(axis = var_943, x = var_942_cast_fp16)[name = string("op_945_cast_fp16")]; |
| string var_949_equation_0 = const()[name = string("op_949_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_949_cast_fp16 = einsum(equation = var_949_equation_0, values = (var_746_cast_fp16_7, var_945_cast_fp16))[name = string("op_949_cast_fp16")]; |
| string scores_equation_0 = const()[name = string("scores_equation_0"), val = string("bkhc,bchq->bkhq")]; |
| tensor<fp16, [1, 1025, 1, 1]> scores_cast_fp16 = einsum(equation = scores_equation_0, values = (var_735_cast_fp16_7, var_713_15))[name = string("scores_cast_fp16")]; |
| tensor<fp16, [1, 1025, 1, 1]> var_955_cast_fp16 = add(x = scores_cast_fp16, y = causal_mask)[name = string("op_955_cast_fp16")]; |
| int32 var_956 = const()[name = string("op_956"), val = int32(1)]; |
| tensor<fp16, [1, 1025, 1, 1]> var_958_cast_fp16 = softmax(axis = var_956, x = var_955_cast_fp16)[name = string("op_958_cast_fp16")]; |
| string var_962_equation_0 = const()[name = string("op_962_equation_0"), val = string("bchk,bkhq->bchq")]; |
| tensor<fp16, [1, 64, 1, 1]> var_962_cast_fp16 = einsum(equation = var_962_equation_0, values = (var_746_cast_fp16_7, var_958_cast_fp16))[name = string("op_962_cast_fp16")]; |
| int32 var_964 = const()[name = string("op_964"), val = int32(1)]; |
| bool input_37_interleave_0 = const()[name = string("input_37_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1024, 1, 1]> input_37_cast_fp16 = concat(axis = var_964, interleave = input_37_interleave_0, values = (var_767_cast_fp16, var_780_cast_fp16, var_793_cast_fp16, var_806_cast_fp16, var_819_cast_fp16, var_832_cast_fp16, var_845_cast_fp16, var_858_cast_fp16, var_871_cast_fp16, var_884_cast_fp16, var_897_cast_fp16, var_910_cast_fp16, var_923_cast_fp16, var_936_cast_fp16, var_949_cast_fp16, var_962_cast_fp16))[name = string("input_37_cast_fp16")]; |
| string out_pad_type_0 = const()[name = string("out_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> out_strides_0 = const()[name = string("out_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> out_pad_0 = const()[name = string("out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> out_dilations_0 = const()[name = string("out_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 out_groups_0 = const()[name = string("out_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 1024, 1, 1]> layers_2_self_attn_out_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68956160))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69480512))))[name = string("layers_2_self_attn_out_proj_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> out_cast_fp16 = conv(dilations = out_dilations_0, groups = out_groups_0, pad = out_pad_0, pad_type = out_pad_type_0, strides = out_strides_0, weight = layers_2_self_attn_out_proj_weight_promoted_to_fp16_palettized, x = input_37_cast_fp16)[name = string("out_cast_fp16")]; |
| tensor<int32, [1]> var_978_axes_0 = const()[name = string("op_978_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_978_cast_fp16 = squeeze(axes = var_978_axes_0, x = out_cast_fp16)[name = string("op_978_cast_fp16")]; |
| tensor<int32, [3]> var_982 = const()[name = string("op_982"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> op_out_5_cast_fp16 = transpose(perm = var_982, x = var_978_cast_fp16)[name = string("transpose_6")]; |
| tensor<fp16, [1, 1, 1024]> x_15_cast_fp16 = add(x = x_9_cast_fp16, y = op_out_5_cast_fp16)[name = string("x_15_cast_fp16")]; |
| fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_986_cast_fp16 = mul(x = x_15_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_986_cast_fp16")]; |
| int32 var_988 = const()[name = string("op_988"), val = int32(-1)]; |
| bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_39_cast_fp16 = concat(axis = var_988, interleave = input_39_interleave_0, values = (x_15_cast_fp16, var_986_cast_fp16))[name = string("input_39_cast_fp16")]; |
| tensor<int32, [1]> normed_19_axes_0 = const()[name = string("normed_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_994_to_fp16 = const()[name = string("op_994_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_19_cast_fp16 = layer_norm(axes = normed_19_axes_0, epsilon = var_994_to_fp16, x = input_39_cast_fp16)[name = string("normed_19_cast_fp16")]; |
| tensor<int32, [2]> var_997_split_sizes_0 = const()[name = string("op_997_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_997_axis_0 = const()[name = string("op_997_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_997_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_997_cast_fp16_1 = split(axis = var_997_axis_0, split_sizes = var_997_split_sizes_0, x = normed_19_cast_fp16)[name = string("op_997_cast_fp16")]; |
| tensor<fp16, [1024]> layers_2_ffn_norm_weight_promoted_to_fp16 = const()[name = string("layers_2_ffn_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69481600)))]; |
| tensor<fp16, [1, 1, 1024]> normed_21_cast_fp16 = mul(x = var_997_cast_fp16_0, y = layers_2_ffn_norm_weight_promoted_to_fp16)[name = string("normed_21_cast_fp16")]; |
| tensor<int32, [3]> var_1003 = const()[name = string("op_1003"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_1006_axes_0 = const()[name = string("op_1006_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1004_cast_fp16 = transpose(perm = var_1003, x = normed_21_cast_fp16)[name = string("transpose_5")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_1006_cast_fp16 = expand_dims(axes = var_1006_axes_0, x = var_1004_cast_fp16)[name = string("op_1006_cast_fp16")]; |
| string input_43_pad_type_0 = const()[name = string("input_43_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_43_strides_0 = const()[name = string("input_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_43_pad_0 = const()[name = string("input_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_43_dilations_0 = const()[name = string("input_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_43_groups_0 = const()[name = string("input_43_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> input_43 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_2_feed_forward_w1_weight_palettized, x = var_1006_cast_fp16)[name = string("input_43")]; |
| string b_5_pad_type_0 = const()[name = string("b_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> b_5_strides_0 = const()[name = string("b_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> b_5_pad_0 = const()[name = string("b_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> b_5_dilations_0 = const()[name = string("b_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 b_5_groups_0 = const()[name = string("b_5_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> b_5 = conv(dilations = b_5_dilations_0, groups = b_5_groups_0, pad = b_5_pad_0, pad_type = b_5_pad_type_0, strides = b_5_strides_0, weight = layers_2_feed_forward_w3_weight_palettized, x = var_1006_cast_fp16)[name = string("b_5")]; |
| tensor<fp16, [1, 4608, 1, 1]> var_1034 = silu(x = input_43)[name = string("op_1034")]; |
| tensor<fp16, [1, 4608, 1, 1]> input_45 = mul(x = var_1034, y = b_5)[name = string("input_45")]; |
| string mlp_9_pad_type_0 = const()[name = string("mlp_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> mlp_9_strides_0 = const()[name = string("mlp_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> mlp_9_pad_0 = const()[name = string("mlp_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> mlp_9_dilations_0 = const()[name = string("mlp_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 mlp_9_groups_0 = const()[name = string("mlp_9_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> mlp_9 = conv(dilations = mlp_9_dilations_0, groups = mlp_9_groups_0, pad = mlp_9_pad_0, pad_type = mlp_9_pad_type_0, strides = mlp_9_strides_0, weight = layers_2_feed_forward_w2_weight_palettized, x = input_45)[name = string("mlp_9")]; |
| tensor<int32, [1]> var_1048_axes_0 = const()[name = string("op_1048_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1048 = squeeze(axes = var_1048_axes_0, x = mlp_9)[name = string("op_1048")]; |
| tensor<int32, [3]> var_1052 = const()[name = string("op_1052"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> mlp_11 = transpose(perm = var_1052, x = var_1048)[name = string("transpose_4")]; |
| tensor<fp16, [1, 1, 1024]> x_17_cast_fp16 = add(x = x_15_cast_fp16, y = mlp_11)[name = string("x_17_cast_fp16")]; |
| fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_1056_cast_fp16 = mul(x = x_17_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_1056_cast_fp16")]; |
| int32 var_1058 = const()[name = string("op_1058"), val = int32(-1)]; |
| bool input_47_interleave_0 = const()[name = string("input_47_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_47_cast_fp16 = concat(axis = var_1058, interleave = input_47_interleave_0, values = (x_17_cast_fp16, var_1056_cast_fp16))[name = string("input_47_cast_fp16")]; |
| tensor<int32, [1]> normed_23_axes_0 = const()[name = string("normed_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_1064_to_fp16 = const()[name = string("op_1064_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_23_cast_fp16 = layer_norm(axes = normed_23_axes_0, epsilon = var_1064_to_fp16, x = input_47_cast_fp16)[name = string("normed_23_cast_fp16")]; |
| tensor<int32, [2]> var_1067_split_sizes_0 = const()[name = string("op_1067_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1067_axis_0 = const()[name = string("op_1067_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_1067_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_1067_cast_fp16_1 = split(axis = var_1067_axis_0, split_sizes = var_1067_split_sizes_0, x = normed_23_cast_fp16)[name = string("op_1067_cast_fp16")]; |
| tensor<fp16, [1024]> layers_3_operator_norm_weight_promoted_to_fp16 = const()[name = string("layers_3_operator_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69483712)))]; |
| tensor<fp16, [1, 1, 1024]> hidden_states_cast_fp16 = mul(x = var_1067_cast_fp16_0, y = layers_3_operator_norm_weight_promoted_to_fp16)[name = string("hidden_states_cast_fp16")]; |
| tensor<int32, [3]> var_1073 = const()[name = string("op_1073"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_1076_axes_0 = const()[name = string("op_1076_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1074_cast_fp16 = transpose(perm = var_1073, x = hidden_states_cast_fp16)[name = string("transpose_3")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_1076_cast_fp16 = expand_dims(axes = var_1076_axes_0, x = var_1074_cast_fp16)[name = string("op_1076_cast_fp16")]; |
| string BCx_pad_type_0 = const()[name = string("BCx_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> BCx_strides_0 = const()[name = string("BCx_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> BCx_pad_0 = const()[name = string("BCx_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> BCx_dilations_0 = const()[name = string("BCx_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 BCx_groups_0 = const()[name = string("BCx_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 3072, 1, 1]> BCx = conv(dilations = BCx_dilations_0, groups = BCx_groups_0, pad = BCx_pad_0, pad_type = BCx_pad_type_0, strides = BCx_strides_0, weight = layers_3_conv_in_proj_weight_palettized, x = var_1076_cast_fp16)[name = string("BCx")]; |
| tensor<int32, [3]> var_1093_split_sizes_0 = const()[name = string("op_1093_split_sizes_0"), val = tensor<int32, [3]>([1024, 1024, 1024])]; |
| int32 var_1093_axis_0 = const()[name = string("op_1093_axis_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> var_1093_0, tensor<fp16, [1, 1024, 1, 1]> var_1093_1, tensor<fp16, [1, 1024, 1, 1]> var_1093_2 = split(axis = var_1093_axis_0, split_sizes = var_1093_split_sizes_0, x = BCx)[name = string("op_1093")]; |
| tensor<fp16, [1, 1024, 1, 1]> Bx = mul(x = var_1093_0, y = var_1093_2)[name = string("Bx")]; |
| tensor<int32, [3]> var_1099_begin_0 = const()[name = string("op_1099_begin_0"), val = tensor<int32, [3]>([2, 0, 0])]; |
| tensor<int32, [3]> var_1099_end_0 = const()[name = string("op_1099_end_0"), val = tensor<int32, [3]>([3, 1024, 3])]; |
| tensor<bool, [3]> var_1099_end_mask_0 = const()[name = string("op_1099_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> var_1099_squeeze_mask_0 = const()[name = string("op_1099_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp16, [1024, 3]> var_1099_cast_fp16 = slice_by_index(begin = var_1099_begin_0, end = var_1099_end_0, end_mask = var_1099_end_mask_0, squeeze_mask = var_1099_squeeze_mask_0, x = conv_state_in)[name = string("op_1099_cast_fp16")]; |
| tensor<int32, [1]> var_1101_axes_0 = const()[name = string("op_1101_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1, 1024, 3]> var_1101_cast_fp16 = expand_dims(axes = var_1101_axes_0, x = var_1099_cast_fp16)[name = string("op_1101_cast_fp16")]; |
| tensor<int32, [1]> slot_axes_0 = const()[name = string("slot_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1, 3]> slot_cast_fp16 = expand_dims(axes = slot_axes_0, x = var_1101_cast_fp16)[name = string("slot_cast_fp16")]; |
| tensor<int32, [4]> live_tail_begin_0 = const()[name = string("live_tail_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1])]; |
| tensor<int32, [4]> live_tail_end_0 = const()[name = string("live_tail_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1])]; |
| tensor<bool, [4]> live_tail_end_mask_0 = const()[name = string("live_tail_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])]; |
| tensor<fp16, [1, 1024, 1, 2]> live_tail_cast_fp16 = slice_by_index(begin = live_tail_begin_0, end = live_tail_end_0, end_mask = live_tail_end_mask_0, x = slot_cast_fp16)[name = string("live_tail_cast_fp16")]; |
| int32 var_1110 = const()[name = string("op_1110"), val = int32(-1)]; |
| bool new_state_interleave_0 = const()[name = string("new_state_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1024, 1, 3]> new_state_cast_fp16 = concat(axis = var_1110, interleave = new_state_interleave_0, values = (live_tail_cast_fp16, Bx))[name = string("new_state_cast_fp16")]; |
| tensor<int32, [1]> var_1113_axes_0 = const()[name = string("op_1113_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1024, 1, 3]> var_1113_cast_fp16 = squeeze(axes = var_1113_axes_0, x = new_state_cast_fp16)[name = string("op_1113_cast_fp16")]; |
| tensor<int32, [1]> new_slot_axes_0 = const()[name = string("new_slot_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp16, [1024, 3]> new_slot_cast_fp16 = squeeze(axes = new_slot_axes_0, x = var_1113_cast_fp16)[name = string("new_slot_cast_fp16")]; |
| string conv_out_pad_type_0 = const()[name = string("conv_out_pad_type_0"), val = string("valid")]; |
| int32 conv_out_groups_0 = const()[name = string("conv_out_groups_0"), val = int32(1024)]; |
| tensor<int32, [2]> conv_out_strides_0 = const()[name = string("conv_out_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> conv_out_pad_0 = const()[name = string("conv_out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> conv_out_dilations_0 = const()[name = string("conv_out_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<fp16, [1024, 1, 1, 3]> layers_3_conv_conv_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1, 1, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69485824))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69487424))))[name = string("layers_3_conv_conv_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> conv_out_cast_fp16 = conv(dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = layers_3_conv_conv_weight_promoted_to_fp16_palettized, x = new_state_cast_fp16)[name = string("conv_out_cast_fp16")]; |
| tensor<fp16, [1, 1024, 1, 1]> input_51_cast_fp16 = mul(x = var_1093_1, y = conv_out_cast_fp16)[name = string("input_51_cast_fp16")]; |
| string y_pad_type_0 = const()[name = string("y_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> y_strides_0 = const()[name = string("y_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> y_pad_0 = const()[name = string("y_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> y_dilations_0 = const()[name = string("y_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 y_groups_0 = const()[name = string("y_groups_0"), val = int32(1)]; |
| tensor<fp16, [1024, 1024, 1, 1]> layers_3_conv_out_proj_weight_promoted_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint4, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69488512))), lut = tensor<fp16, [32, 1, 1, 1, 16, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70012864))))[name = string("layers_3_conv_out_proj_weight_promoted_to_fp16_palettized")]; |
| tensor<fp16, [1, 1024, 1, 1]> y_cast_fp16 = conv(dilations = y_dilations_0, groups = y_groups_0, pad = y_pad_0, pad_type = y_pad_type_0, strides = y_strides_0, weight = layers_3_conv_out_proj_weight_promoted_to_fp16_palettized, x = input_51_cast_fp16)[name = string("y_cast_fp16")]; |
| tensor<int32, [1]> var_1141_axes_0 = const()[name = string("op_1141_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1141_cast_fp16 = squeeze(axes = var_1141_axes_0, x = y_cast_fp16)[name = string("op_1141_cast_fp16")]; |
| tensor<int32, [3]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> op_out_cast_fp16 = transpose(perm = var_1145, x = var_1141_cast_fp16)[name = string("transpose_2")]; |
| tensor<fp16, [1, 1, 1024]> x_cast_fp16 = add(x = x_17_cast_fp16, y = op_out_cast_fp16)[name = string("x_cast_fp16")]; |
| fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)]; |
| tensor<fp16, [1, 1, 1024]> var_1149_cast_fp16 = mul(x = x_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_1149_cast_fp16")]; |
| int32 var_1151 = const()[name = string("op_1151"), val = int32(-1)]; |
| bool input_53_interleave_0 = const()[name = string("input_53_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 1, 2048]> input_53_cast_fp16 = concat(axis = var_1151, interleave = input_53_interleave_0, values = (x_cast_fp16, var_1149_cast_fp16))[name = string("input_53_cast_fp16")]; |
| tensor<int32, [1]> normed_25_axes_0 = const()[name = string("normed_25_axes_0"), val = tensor<int32, [1]>([-1])]; |
| fp16 var_1157_to_fp16 = const()[name = string("op_1157_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1, 2048]> normed_25_cast_fp16 = layer_norm(axes = normed_25_axes_0, epsilon = var_1157_to_fp16, x = input_53_cast_fp16)[name = string("normed_25_cast_fp16")]; |
| tensor<int32, [2]> var_1160_split_sizes_0 = const()[name = string("op_1160_split_sizes_0"), val = tensor<int32, [2]>([1024, 1024])]; |
| int32 var_1160_axis_0 = const()[name = string("op_1160_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 1, 1024]> var_1160_cast_fp16_0, tensor<fp16, [1, 1, 1024]> var_1160_cast_fp16_1 = split(axis = var_1160_axis_0, split_sizes = var_1160_split_sizes_0, x = normed_25_cast_fp16)[name = string("op_1160_cast_fp16")]; |
| tensor<fp16, [1024]> layers_3_ffn_norm_weight_promoted_to_fp16 = const()[name = string("layers_3_ffn_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70013952)))]; |
| tensor<fp16, [1, 1, 1024]> normed_cast_fp16 = mul(x = var_1160_cast_fp16_0, y = layers_3_ffn_norm_weight_promoted_to_fp16)[name = string("normed_cast_fp16")]; |
| tensor<int32, [3]> var_1166 = const()[name = string("op_1166"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_1169_axes_0 = const()[name = string("op_1169_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1167_cast_fp16 = transpose(perm = var_1166, x = normed_cast_fp16)[name = string("transpose_1")]; |
| tensor<fp16, [1, 1024, 1, 1]> var_1169_cast_fp16 = expand_dims(axes = var_1169_axes_0, x = var_1167_cast_fp16)[name = string("op_1169_cast_fp16")]; |
| string input_57_pad_type_0 = const()[name = string("input_57_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> input_57_strides_0 = const()[name = string("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> input_57_pad_0 = const()[name = string("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> input_57_dilations_0 = const()[name = string("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 input_57_groups_0 = const()[name = string("input_57_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> input_57 = conv(dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_3_feed_forward_w1_weight_palettized, x = var_1169_cast_fp16)[name = string("input_57")]; |
| string b_pad_type_0 = const()[name = string("b_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> b_strides_0 = const()[name = string("b_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> b_pad_0 = const()[name = string("b_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> b_dilations_0 = const()[name = string("b_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 b_groups_0 = const()[name = string("b_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 4608, 1, 1]> b = conv(dilations = b_dilations_0, groups = b_groups_0, pad = b_pad_0, pad_type = b_pad_type_0, strides = b_strides_0, weight = layers_3_feed_forward_w3_weight_palettized, x = var_1169_cast_fp16)[name = string("b")]; |
| tensor<fp16, [1, 4608, 1, 1]> var_1197 = silu(x = input_57)[name = string("op_1197")]; |
| tensor<fp16, [1, 4608, 1, 1]> input = mul(x = var_1197, y = b)[name = string("input")]; |
| string mlp_13_pad_type_0 = const()[name = string("mlp_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [2]> mlp_13_strides_0 = const()[name = string("mlp_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [4]> mlp_13_pad_0 = const()[name = string("mlp_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [2]> mlp_13_dilations_0 = const()[name = string("mlp_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 mlp_13_groups_0 = const()[name = string("mlp_13_groups_0"), val = int32(1)]; |
| tensor<fp16, [1, 1024, 1, 1]> mlp_13 = conv(dilations = mlp_13_dilations_0, groups = mlp_13_groups_0, pad = mlp_13_pad_0, pad_type = mlp_13_pad_type_0, strides = mlp_13_strides_0, weight = layers_3_feed_forward_w2_weight_palettized, x = input)[name = string("mlp_13")]; |
| tensor<int32, [1]> var_1211_axes_0 = const()[name = string("op_1211_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp16, [1, 1024, 1]> var_1211 = squeeze(axes = var_1211_axes_0, x = mlp_13)[name = string("op_1211")]; |
| tensor<int32, [3]> var_1215 = const()[name = string("op_1215"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp16, [1, 1, 1024]> mlp = transpose(perm = var_1215, x = var_1211)[name = string("transpose_0")]; |
| tensor<fp16, [1, 1, 1024]> hidden_out = add(x = x_cast_fp16, y = mlp)[name = string("op_1218_cast_fp16")]; |
| int32 var_1221_axis_0 = const()[name = string("op_1221_axis_0"), val = int32(0)]; |
| tensor<fp16, [3, 1024, 3]> conv_state_out = stack(axis = var_1221_axis_0, values = (var_237_cast_fp16, var_400_cast_fp16, new_slot_cast_fp16))[name = string("op_1221_cast_fp16")]; |
| tensor<int32, [1]> var_1224_axes_0 = const()[name = string("op_1224_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1, 1, 512, 1, 1]> var_1224 = expand_dims(axes = var_1224_axes_0, x = k_slice)[name = string("op_1224")]; |
| tensor<int32, [1]> var_1227_axes_0 = const()[name = string("op_1227_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp16, [1, 1, 512, 1, 1]> var_1227 = expand_dims(axes = var_1227_axes_0, x = var_596)[name = string("op_1227")]; |
| int32 var_1229 = const()[name = string("op_1229"), val = int32(0)]; |
| bool var_1230_interleave_0 = const()[name = string("op_1230_interleave_0"), val = bool(false)]; |
| tensor<fp16, [2, 1, 512, 1, 1]> kv_slice_out = concat(axis = var_1229, interleave = var_1230_interleave_0, values = (var_1224, var_1227))[name = string("op_1230")]; |
| tensor<fp16, [1, 1, 1024, 1]> update_mask_tmp = identity(x = update_mask)[name = string("update_mask_tmp")]; |
| } -> (hidden_out, kv_slice_out, conv_state_out); |
| } |
