diff --git "a/small/TextDecoder.mlmodelc/model.mil" "b/small/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/small/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,1025 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})] +{ + func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_24_axis_0 = const()[name = tensor("op_24_axis_0"), val = tensor(0)]; + tensor var_24_batch_dims_0 = const()[name = tensor("op_24_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_24_cast_fp16")]; + tensor var_31_axis_0 = const()[name = tensor("op_31_axis_0"), val = tensor(0)]; + tensor var_31_batch_dims_0 = const()[name = tensor("op_31_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_inlier_module_weight_to_fp16 = const()[name = tensor("embed_positions_inlier_module_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; + tensor var_31_cast_fp16 = gather(axis = var_31_axis_0, batch_dims = var_31_batch_dims_0, indices = cache_length, x = embed_positions_inlier_module_weight_to_fp16)[name = tensor("op_31_cast_fp16")]; + tensor var_33_axis_0 = const()[name = tensor("op_33_axis_0"), val = tensor(0)]; + tensor var_33_batch_dims_0 = const()[name = tensor("op_33_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133940800))), name = tensor("embed_positions_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924032))), shape = tensor([448, 1280])]; + tensor var_33_cast_fp16 = gather(axis = var_33_axis_0, batch_dims = var_33_batch_dims_0, indices = cache_length, x = embed_positions_outlier_module_weight_to_fp16_sparsified)[name = tensor("op_33_cast_fp16")]; + tensor var_34_cast_fp16 = add(x = var_31_cast_fp16, y = var_33_cast_fp16)[name = tensor("op_34_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_34_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_48_axes_0 = const()[name = tensor("op_48_axes_0"), val = tensor([2])]; + tensor var_48_cast_fp16 = expand_dims(axes = var_48_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_48_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_48_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280])]; + tensor var_53_axis_0 = const()[name = tensor("op_53_axis_0"), val = tensor(1)]; + tensor var_53_cast_fp16_0, tensor var_53_cast_fp16_1, tensor var_53_cast_fp16_2, tensor var_53_cast_fp16_3 = split(axis = var_53_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_53_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280])]; + tensor var_60_axis_0 = const()[name = tensor("op_60_axis_0"), val = tensor(1)]; + tensor var_60_cast_fp16_0, tensor var_60_cast_fp16_1, tensor var_60_cast_fp16_2, tensor var_60_cast_fp16_3 = split(axis = var_60_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_60_cast_fp16")]; + tensor var_70 = const()[name = tensor("op_70"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_95_to_fp16 = const()[name = tensor("op_95_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_95_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134012544)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134015168)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134017792)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134020416)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor var_117_pad_type_0 = const()[name = tensor("op_117_pad_type_0"), val = tensor("valid")]; + tensor var_117_strides_0 = const()[name = tensor("op_117_strides_0"), val = tensor([1, 1])]; + tensor var_117_pad_0 = const()[name = tensor("op_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_117_dilations_0 = const()[name = tensor("op_117_dilations_0"), val = tensor([1, 1])]; + tensor var_117_groups_0 = const()[name = tensor("op_117_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134023040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135661504))), name = tensor("layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135662080)))]; + tensor var_117_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_117_dilations_0, groups = var_117_groups_0, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_117_strides_0, weight = layers_0_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("op_117_cast_fp16")]; + tensor var_123_pad_type_0 = const()[name = tensor("op_123_pad_type_0"), val = tensor("valid")]; + tensor var_123_strides_0 = const()[name = tensor("op_123_strides_0"), val = tensor([1, 1])]; + tensor var_123_pad_0 = const()[name = tensor("op_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_123_dilations_0 = const()[name = tensor("op_123_dilations_0"), val = tensor([1, 1])]; + tensor var_123_groups_0 = const()[name = tensor("op_123_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135737472))), name = tensor("layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135664704))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_123_cast_fp16 = conv(dilations = var_123_dilations_0, groups = var_123_groups_0, pad = var_123_pad_0, pad_type = var_123_pad_type_0, strides = var_123_strides_0, weight = layers_0_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor("op_123_cast_fp16")]; + tensor query_1_cast_fp16 = add(x = var_117_cast_fp16, y = var_123_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_132_pad_type_0 = const()[name = tensor("op_132_pad_type_0"), val = tensor("valid")]; + tensor var_132_strides_0 = const()[name = tensor("op_132_strides_0"), val = tensor([1, 1])]; + tensor var_132_pad_0 = const()[name = tensor("op_132_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_132_dilations_0 = const()[name = tensor("op_132_dilations_0"), val = tensor([1, 1])]; + tensor var_132_groups_0 = const()[name = tensor("op_132_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135942336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137580800))), name = tensor("layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_132_cast_fp16 = conv(dilations = var_132_dilations_0, groups = var_132_groups_0, pad = var_132_pad_0, pad_type = var_132_pad_type_0, strides = var_132_strides_0, weight = layers_0_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("op_132_cast_fp16")]; + tensor var_138_pad_type_0 = const()[name = tensor("op_138_pad_type_0"), val = tensor("valid")]; + tensor var_138_strides_0 = const()[name = tensor("op_138_strides_0"), val = tensor([1, 1])]; + tensor var_138_pad_0 = const()[name = tensor("op_138_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_138_dilations_0 = const()[name = tensor("op_138_dilations_0"), val = tensor([1, 1])]; + tensor var_138_groups_0 = const()[name = tensor("op_138_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137614080))), name = tensor("layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137581376))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_138_cast_fp16 = conv(dilations = var_138_dilations_0, groups = var_138_groups_0, pad = var_138_pad_0, pad_type = var_138_pad_type_0, strides = var_138_strides_0, weight = layers_0_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor("op_138_cast_fp16")]; + tensor current_key_1_cast_fp16 = add(x = var_132_cast_fp16, y = var_138_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_148_pad_type_0 = const()[name = tensor("op_148_pad_type_0"), val = tensor("valid")]; + tensor var_148_strides_0 = const()[name = tensor("op_148_strides_0"), val = tensor([1, 1])]; + tensor var_148_pad_0 = const()[name = tensor("op_148_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_148_dilations_0 = const()[name = tensor("op_148_dilations_0"), val = tensor([1, 1])]; + tensor var_148_groups_0 = const()[name = tensor("op_148_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137818944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139457408))), name = tensor("layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139457984)))]; + tensor var_148_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_148_dilations_0, groups = var_148_groups_0, pad = var_148_pad_0, pad_type = var_148_pad_type_0, strides = var_148_strides_0, weight = layers_0_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("op_148_cast_fp16")]; + tensor var_154_pad_type_0 = const()[name = tensor("op_154_pad_type_0"), val = tensor("valid")]; + tensor var_154_strides_0 = const()[name = tensor("op_154_strides_0"), val = tensor([1, 1])]; + tensor var_154_pad_0 = const()[name = tensor("op_154_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_154_dilations_0 = const()[name = tensor("op_154_dilations_0"), val = tensor([1, 1])]; + tensor var_154_groups_0 = const()[name = tensor("op_154_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139503488))), name = tensor("layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139460608))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_154_cast_fp16 = conv(dilations = var_154_dilations_0, groups = var_154_groups_0, pad = var_154_pad_0, pad_type = var_154_pad_type_0, strides = var_154_strides_0, weight = layers_0_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor current_value_1_cast_fp16 = add(x = var_148_cast_fp16, y = var_154_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_157_axes_0 = const()[name = tensor("op_157_axes_0"), val = tensor([1])]; + tensor var_157_cast_fp16 = expand_dims(axes = var_157_axes_0, x = kv_cache_update_mask)[name = tensor("op_157_cast_fp16")]; + tensor var_158_axes_0 = const()[name = tensor("op_158_axes_0"), val = tensor([2])]; + tensor var_158_cast_fp16 = expand_dims(axes = var_158_axes_0, x = var_157_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_71_to_fp16 = const()[name = tensor("op_71_to_fp16"), val = tensor(0x1p+0)]; + tensor var_160_cast_fp16 = sub(x = var_71_to_fp16, y = var_158_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = var_53_cast_fp16_0, y = var_160_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor var_162_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = var_60_cast_fp16_0, y = var_160_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor var_166_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_165_cast_fp16, y = var_166_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_170, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_172_to_fp16 = const()[name = tensor("op_172_to_fp16"), val = tensor(0x1p-3)]; + tensor var_173_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_172_to_fp16)[name = tensor("op_173_cast_fp16")]; + tensor var_176 = const()[name = tensor("op_176"), val = tensor([1, 20, 64, 448])]; + tensor var_177_cast_fp16 = reshape(shape = var_176, x = key_1_cast_fp16)[name = tensor("op_177_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_173_cast_fp16, y = var_177_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_181_axes_0 = const()[name = tensor("op_181_axes_0"), val = tensor([1])]; + tensor var_181_cast_fp16 = expand_dims(axes = var_181_axes_0, x = decoder_key_padding_mask)[name = tensor("op_181_cast_fp16")]; + tensor var_182_axes_0 = const()[name = tensor("op_182_axes_0"), val = tensor([2])]; + tensor var_182_cast_fp16 = expand_dims(axes = var_182_axes_0, x = var_181_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_182_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_185_cast_fp16 = softmax(axis = var_70, x = mh_w_3_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 20, 64, 448])]; + tensor var_187_cast_fp16 = reshape(shape = var_186, x = value_1_cast_fp16)[name = tensor("op_187_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_187_cast_fp16, y = var_185_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_190 = const()[name = tensor("op_190"), val = tensor([1, 1280, 1, 1])]; + tensor input_1_cast_fp16 = reshape(shape = var_190, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_200_pad_type_0 = const()[name = tensor("op_200_pad_type_0"), val = tensor("valid")]; + tensor var_200_strides_0 = const()[name = tensor("op_200_strides_0"), val = tensor([1, 1])]; + tensor var_200_pad_0 = const()[name = tensor("op_200_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_200_dilations_0 = const()[name = tensor("op_200_dilations_0"), val = tensor([1, 1])]; + tensor var_200_groups_0 = const()[name = tensor("op_200_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139708352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141346816))), name = tensor("layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141347392)))]; + tensor var_200_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_200_dilations_0, groups = var_200_groups_0, pad = var_200_pad_0, pad_type = var_200_pad_type_0, strides = var_200_strides_0, weight = layers_0_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_206_pad_type_0 = const()[name = tensor("op_206_pad_type_0"), val = tensor("valid")]; + tensor var_206_strides_0 = const()[name = tensor("op_206_strides_0"), val = tensor([1, 1])]; + tensor var_206_pad_0 = const()[name = tensor("op_206_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_206_dilations_0 = const()[name = tensor("op_206_dilations_0"), val = tensor([1, 1])]; + tensor var_206_groups_0 = const()[name = tensor("op_206_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141407744))), name = tensor("layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141350016))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_206_cast_fp16 = conv(dilations = var_206_dilations_0, groups = var_206_groups_0, pad = var_206_pad_0, pad_type = var_206_pad_type_0, strides = var_206_strides_0, weight = layers_0_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_1_cast_fp16)[name = tensor("op_206_cast_fp16")]; + tensor obj_7_cast_fp16 = add(x = var_200_cast_fp16, y = var_206_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_221_to_fp16 = const()[name = tensor("op_221_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_221_to_fp16, x = inputs_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141612608)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141615232)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_243_pad_type_0 = const()[name = tensor("op_243_pad_type_0"), val = tensor("valid")]; + tensor var_243_strides_0 = const()[name = tensor("op_243_strides_0"), val = tensor([1, 1])]; + tensor var_243_pad_0 = const()[name = tensor("op_243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_243_dilations_0 = const()[name = tensor("op_243_dilations_0"), val = tensor([1, 1])]; + tensor var_243_groups_0 = const()[name = tensor("op_243_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141617856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143256320))), name = tensor("layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143256896)))]; + tensor var_243_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_243_dilations_0, groups = var_243_groups_0, pad = var_243_pad_0, pad_type = var_243_pad_type_0, strides = var_243_strides_0, weight = layers_0_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("op_243_cast_fp16")]; + tensor var_249_pad_type_0 = const()[name = tensor("op_249_pad_type_0"), val = tensor("valid")]; + tensor var_249_strides_0 = const()[name = tensor("op_249_strides_0"), val = tensor([1, 1])]; + tensor var_249_pad_0 = const()[name = tensor("op_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_249_dilations_0 = const()[name = tensor("op_249_dilations_0"), val = tensor([1, 1])]; + tensor var_249_groups_0 = const()[name = tensor("op_249_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143284992))), name = tensor("layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143259520))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_249_cast_fp16 = conv(dilations = var_249_dilations_0, groups = var_249_groups_0, pad = var_249_pad_0, pad_type = var_249_pad_type_0, strides = var_249_strides_0, weight = layers_0_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_9_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor query_3_cast_fp16 = add(x = var_243_cast_fp16, y = var_249_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_258_pad_type_0 = const()[name = tensor("op_258_pad_type_0"), val = tensor("valid")]; + tensor var_258_strides_0 = const()[name = tensor("op_258_strides_0"), val = tensor([1, 1])]; + tensor var_258_pad_0 = const()[name = tensor("op_258_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_258_dilations_0 = const()[name = tensor("op_258_dilations_0"), val = tensor([1, 1])]; + tensor var_258_groups_0 = const()[name = tensor("op_258_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143489856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145128320))), name = tensor("layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_258_cast_fp16 = conv(dilations = var_258_dilations_0, groups = var_258_groups_0, pad = var_258_pad_0, pad_type = var_258_pad_type_0, strides = var_258_strides_0, weight = layers_0_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_258_cast_fp16")]; + tensor var_264_pad_type_0 = const()[name = tensor("op_264_pad_type_0"), val = tensor("valid")]; + tensor var_264_strides_0 = const()[name = tensor("op_264_strides_0"), val = tensor([1, 1])]; + tensor var_264_pad_0 = const()[name = tensor("op_264_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_264_dilations_0 = const()[name = tensor("op_264_dilations_0"), val = tensor([1, 1])]; + tensor var_264_groups_0 = const()[name = tensor("op_264_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145188864))), name = tensor("layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145128896))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_264_cast_fp16 = conv(dilations = var_264_dilations_0, groups = var_264_groups_0, pad = var_264_pad_0, pad_type = var_264_pad_type_0, strides = var_264_strides_0, weight = layers_0_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_264_cast_fp16")]; + tensor key_3_cast_fp16 = add(x = var_258_cast_fp16, y = var_264_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor var_274_pad_type_0 = const()[name = tensor("op_274_pad_type_0"), val = tensor("valid")]; + tensor var_274_strides_0 = const()[name = tensor("op_274_strides_0"), val = tensor([1, 1])]; + tensor var_274_pad_0 = const()[name = tensor("op_274_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_274_dilations_0 = const()[name = tensor("op_274_dilations_0"), val = tensor([1, 1])]; + tensor var_274_groups_0 = const()[name = tensor("op_274_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145393728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147032192))), name = tensor("layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147032768)))]; + tensor var_274_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_274_dilations_0, groups = var_274_groups_0, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_274_strides_0, weight = layers_0_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_274_cast_fp16")]; + tensor var_280_pad_type_0 = const()[name = tensor("op_280_pad_type_0"), val = tensor("valid")]; + tensor var_280_strides_0 = const()[name = tensor("op_280_strides_0"), val = tensor([1, 1])]; + tensor var_280_pad_0 = const()[name = tensor("op_280_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_280_dilations_0 = const()[name = tensor("op_280_dilations_0"), val = tensor([1, 1])]; + tensor var_280_groups_0 = const()[name = tensor("op_280_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147046656))), name = tensor("layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147035392))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_280_cast_fp16 = conv(dilations = var_280_dilations_0, groups = var_280_groups_0, pad = var_280_pad_0, pad_type = var_280_pad_type_0, strides = var_280_strides_0, weight = layers_0_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_280_cast_fp16")]; + tensor value_3_cast_fp16 = add(x = var_274_cast_fp16, y = var_280_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_284, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_286_to_fp16 = const()[name = tensor("op_286_to_fp16"), val = tensor(0x1p-3)]; + tensor var_287_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_286_to_fp16)[name = tensor("op_287_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 20, 64, 1500])]; + tensor var_291_cast_fp16 = reshape(shape = var_290, x = key_3_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_287_cast_fp16, y = var_291_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_70, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_295 = const()[name = tensor("op_295"), val = tensor([1, 20, 64, 1500])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = value_3_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_296_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_299 = const()[name = tensor("op_299"), val = tensor([1, 1280, 1, 1])]; + tensor input_3_cast_fp16 = reshape(shape = var_299, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_309_pad_type_0 = const()[name = tensor("op_309_pad_type_0"), val = tensor("valid")]; + tensor var_309_strides_0 = const()[name = tensor("op_309_strides_0"), val = tensor([1, 1])]; + tensor var_309_pad_0 = const()[name = tensor("op_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_309_dilations_0 = const()[name = tensor("op_309_dilations_0"), val = tensor([1, 1])]; + tensor var_309_groups_0 = const()[name = tensor("op_309_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147251520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148889984))), name = tensor("layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148890560)))]; + tensor var_309_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_309_dilations_0, groups = var_309_groups_0, pad = var_309_pad_0, pad_type = var_309_pad_type_0, strides = var_309_strides_0, weight = layers_0_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_315_pad_type_0 = const()[name = tensor("op_315_pad_type_0"), val = tensor("valid")]; + tensor var_315_strides_0 = const()[name = tensor("op_315_strides_0"), val = tensor([1, 1])]; + tensor var_315_pad_0 = const()[name = tensor("op_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_315_dilations_0 = const()[name = tensor("op_315_dilations_0"), val = tensor([1, 1])]; + tensor var_315_groups_0 = const()[name = tensor("op_315_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148905152))), name = tensor("layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148893184))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_315_cast_fp16 = conv(dilations = var_315_dilations_0, groups = var_315_groups_0, pad = var_315_pad_0, pad_type = var_315_pad_type_0, strides = var_315_strides_0, weight = layers_0_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_3_cast_fp16)[name = tensor("op_315_cast_fp16")]; + tensor obj_11_cast_fp16 = add(x = var_309_cast_fp16, y = var_315_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_326_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149110016)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149112640)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_344_pad_type_0 = const()[name = tensor("op_344_pad_type_0"), val = tensor("valid")]; + tensor var_344_strides_0 = const()[name = tensor("op_344_strides_0"), val = tensor([1, 1])]; + tensor var_344_pad_0 = const()[name = tensor("op_344_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_344_dilations_0 = const()[name = tensor("op_344_dilations_0"), val = tensor([1, 1])]; + tensor var_344_groups_0 = const()[name = tensor("op_344_groups_0"), val = tensor(1)]; + tensor layers_0_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149115264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155668928))), name = tensor("layers_0_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_0_fc1_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_fc1_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155669504)))]; + tensor var_344_cast_fp16 = conv(bias = layers_0_fc1_inlier_module_bias_to_fp16, dilations = var_344_dilations_0, groups = var_344_groups_0, pad = var_344_pad_0, pad_type = var_344_pad_type_0, strides = var_344_strides_0, weight = layers_0_fc1_inlier_module_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_350_pad_type_0 = const()[name = tensor("op_350_pad_type_0"), val = tensor("valid")]; + tensor var_350_strides_0 = const()[name = tensor("op_350_strides_0"), val = tensor([1, 1])]; + tensor var_350_pad_0 = const()[name = tensor("op_350_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_350_dilations_0 = const()[name = tensor("op_350_dilations_0"), val = tensor([1, 1])]; + tensor var_350_groups_0 = const()[name = tensor("op_350_groups_0"), val = tensor(1)]; + tensor layers_0_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155781696))), name = tensor("layers_0_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155679808))), shape = tensor([5120, 1280, 1, 1])]; + tensor var_350_cast_fp16 = conv(dilations = var_350_dilations_0, groups = var_350_groups_0, pad = var_350_pad_0, pad_type = var_350_pad_type_0, strides = var_350_strides_0, weight = layers_0_fc1_outlier_module_weight_to_fp16_sparsified, x = input_5_cast_fp16)[name = tensor("op_350_cast_fp16")]; + tensor input_7_cast_fp16 = add(x = var_344_cast_fp16, y = var_350_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_361_pad_type_0 = const()[name = tensor("op_361_pad_type_0"), val = tensor("valid")]; + tensor var_361_strides_0 = const()[name = tensor("op_361_strides_0"), val = tensor([1, 1])]; + tensor var_361_pad_0 = const()[name = tensor("op_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_361_dilations_0 = const()[name = tensor("op_361_dilations_0"), val = tensor([1, 1])]; + tensor var_361_groups_0 = const()[name = tensor("op_361_groups_0"), val = tensor(1)]; + tensor layers_0_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156600960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163154624))), name = tensor("layers_0_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_0_fc2_inlier_module_bias_to_fp16 = const()[name = tensor("layers_0_fc2_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163155200)))]; + tensor var_361_cast_fp16 = conv(bias = layers_0_fc2_inlier_module_bias_to_fp16, dilations = var_361_dilations_0, groups = var_361_groups_0, pad = var_361_pad_0, pad_type = var_361_pad_type_0, strides = var_361_strides_0, weight = layers_0_fc2_inlier_module_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor("op_361_cast_fp16")]; + tensor var_367_pad_type_0 = const()[name = tensor("op_367_pad_type_0"), val = tensor("valid")]; + tensor var_367_strides_0 = const()[name = tensor("op_367_strides_0"), val = tensor([1, 1])]; + tensor var_367_pad_0 = const()[name = tensor("op_367_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_367_dilations_0 = const()[name = tensor("op_367_dilations_0"), val = tensor([1, 1])]; + tensor var_367_groups_0 = const()[name = tensor("op_367_groups_0"), val = tensor(1)]; + tensor layers_0_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163336192))), name = tensor("layers_0_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163157824))), shape = tensor([1280, 5120, 1, 1])]; + tensor var_367_cast_fp16 = conv(dilations = var_367_dilations_0, groups = var_367_groups_0, pad = var_367_pad_0, pad_type = var_367_pad_type_0, strides = var_367_strides_0, weight = layers_0_fc2_outlier_module_weight_to_fp16_sparsified, x = input_9_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor hidden_states_3_cast_fp16 = add(x = var_361_cast_fp16, y = var_367_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_404_to_fp16 = const()[name = tensor("op_404_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_404_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164155456)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164158080)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor var_426_pad_type_0 = const()[name = tensor("op_426_pad_type_0"), val = tensor("valid")]; + tensor var_426_strides_0 = const()[name = tensor("op_426_strides_0"), val = tensor([1, 1])]; + tensor var_426_pad_0 = const()[name = tensor("op_426_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_426_dilations_0 = const()[name = tensor("op_426_dilations_0"), val = tensor([1, 1])]; + tensor var_426_groups_0 = const()[name = tensor("op_426_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164160704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165799168))), name = tensor("layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165799744)))]; + tensor var_426_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_426_dilations_0, groups = var_426_groups_0, pad = var_426_pad_0, pad_type = var_426_pad_type_0, strides = var_426_strides_0, weight = layers_1_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("op_426_cast_fp16")]; + tensor var_432_pad_type_0 = const()[name = tensor("op_432_pad_type_0"), val = tensor("valid")]; + tensor var_432_strides_0 = const()[name = tensor("op_432_strides_0"), val = tensor([1, 1])]; + tensor var_432_pad_0 = const()[name = tensor("op_432_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_432_dilations_0 = const()[name = tensor("op_432_dilations_0"), val = tensor([1, 1])]; + tensor var_432_groups_0 = const()[name = tensor("op_432_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165862272))), name = tensor("layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165802368))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_432_cast_fp16 = conv(dilations = var_432_dilations_0, groups = var_432_groups_0, pad = var_432_pad_0, pad_type = var_432_pad_type_0, strides = var_432_strides_0, weight = layers_1_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor("op_432_cast_fp16")]; + tensor query_5_cast_fp16 = add(x = var_426_cast_fp16, y = var_432_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_441_pad_type_0 = const()[name = tensor("op_441_pad_type_0"), val = tensor("valid")]; + tensor var_441_strides_0 = const()[name = tensor("op_441_strides_0"), val = tensor([1, 1])]; + tensor var_441_pad_0 = const()[name = tensor("op_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_441_dilations_0 = const()[name = tensor("op_441_dilations_0"), val = tensor([1, 1])]; + tensor var_441_groups_0 = const()[name = tensor("op_441_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166067136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167705600))), name = tensor("layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_441_cast_fp16 = conv(dilations = var_441_dilations_0, groups = var_441_groups_0, pad = var_441_pad_0, pad_type = var_441_pad_type_0, strides = var_441_strides_0, weight = layers_1_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("op_441_cast_fp16")]; + tensor var_447_pad_type_0 = const()[name = tensor("op_447_pad_type_0"), val = tensor("valid")]; + tensor var_447_strides_0 = const()[name = tensor("op_447_strides_0"), val = tensor([1, 1])]; + tensor var_447_pad_0 = const()[name = tensor("op_447_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_447_dilations_0 = const()[name = tensor("op_447_dilations_0"), val = tensor([1, 1])]; + tensor var_447_groups_0 = const()[name = tensor("op_447_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167752448))), name = tensor("layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167706176))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_447_cast_fp16 = conv(dilations = var_447_dilations_0, groups = var_447_groups_0, pad = var_447_pad_0, pad_type = var_447_pad_type_0, strides = var_447_strides_0, weight = layers_1_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor current_key_3_cast_fp16 = add(x = var_441_cast_fp16, y = var_447_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_457_pad_type_0 = const()[name = tensor("op_457_pad_type_0"), val = tensor("valid")]; + tensor var_457_strides_0 = const()[name = tensor("op_457_strides_0"), val = tensor([1, 1])]; + tensor var_457_pad_0 = const()[name = tensor("op_457_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_457_dilations_0 = const()[name = tensor("op_457_dilations_0"), val = tensor([1, 1])]; + tensor var_457_groups_0 = const()[name = tensor("op_457_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167957312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169595776))), name = tensor("layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169596352)))]; + tensor var_457_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_457_dilations_0, groups = var_457_groups_0, pad = var_457_pad_0, pad_type = var_457_pad_type_0, strides = var_457_strides_0, weight = layers_1_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_pad_type_0 = const()[name = tensor("op_463_pad_type_0"), val = tensor("valid")]; + tensor var_463_strides_0 = const()[name = tensor("op_463_strides_0"), val = tensor([1, 1])]; + tensor var_463_pad_0 = const()[name = tensor("op_463_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_463_dilations_0 = const()[name = tensor("op_463_dilations_0"), val = tensor([1, 1])]; + tensor var_463_groups_0 = const()[name = tensor("op_463_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169621568))), name = tensor("layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169598976))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_463_cast_fp16 = conv(dilations = var_463_dilations_0, groups = var_463_groups_0, pad = var_463_pad_0, pad_type = var_463_pad_type_0, strides = var_463_strides_0, weight = layers_1_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_15_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor current_value_3_cast_fp16 = add(x = var_457_cast_fp16, y = var_463_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = var_53_cast_fp16_1, y = var_160_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_470_cast_fp16, y = var_471_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_474_cast_fp16 = mul(x = var_60_cast_fp16_1, y = var_160_cast_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_475_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_475_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_474_cast_fp16, y = var_475_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_479 = const()[name = tensor("op_479"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_479, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_481_to_fp16 = const()[name = tensor("op_481_to_fp16"), val = tensor(0x1p-3)]; + tensor var_482_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_481_to_fp16)[name = tensor("op_482_cast_fp16")]; + tensor var_485 = const()[name = tensor("op_485"), val = tensor([1, 20, 64, 448])]; + tensor var_486_cast_fp16 = reshape(shape = var_485, x = key_5_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_482_cast_fp16, y = var_486_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_182_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_494_cast_fp16 = softmax(axis = var_379, x = mh_w_9_cast_fp16)[name = tensor("op_494_cast_fp16")]; + tensor var_495 = const()[name = tensor("op_495"), val = tensor([1, 20, 64, 448])]; + tensor var_496_cast_fp16 = reshape(shape = var_495, x = value_5_cast_fp16)[name = tensor("op_496_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_496_cast_fp16, y = var_494_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_499 = const()[name = tensor("op_499"), val = tensor([1, 1280, 1, 1])]; + tensor input_11_cast_fp16 = reshape(shape = var_499, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_509_pad_type_0 = const()[name = tensor("op_509_pad_type_0"), val = tensor("valid")]; + tensor var_509_strides_0 = const()[name = tensor("op_509_strides_0"), val = tensor([1, 1])]; + tensor var_509_pad_0 = const()[name = tensor("op_509_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_509_dilations_0 = const()[name = tensor("op_509_dilations_0"), val = tensor([1, 1])]; + tensor var_509_groups_0 = const()[name = tensor("op_509_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169826432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171464896))), name = tensor("layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171465472)))]; + tensor var_509_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_509_dilations_0, groups = var_509_groups_0, pad = var_509_pad_0, pad_type = var_509_pad_type_0, strides = var_509_strides_0, weight = layers_1_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("op_509_cast_fp16")]; + tensor var_515_pad_type_0 = const()[name = tensor("op_515_pad_type_0"), val = tensor("valid")]; + tensor var_515_strides_0 = const()[name = tensor("op_515_strides_0"), val = tensor([1, 1])]; + tensor var_515_pad_0 = const()[name = tensor("op_515_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_515_dilations_0 = const()[name = tensor("op_515_dilations_0"), val = tensor([1, 1])]; + tensor var_515_groups_0 = const()[name = tensor("op_515_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171492800))), name = tensor("layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171468096))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_515_cast_fp16 = conv(dilations = var_515_dilations_0, groups = var_515_groups_0, pad = var_515_pad_0, pad_type = var_515_pad_type_0, strides = var_515_strides_0, weight = layers_1_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_11_cast_fp16)[name = tensor("op_515_cast_fp16")]; + tensor obj_21_cast_fp16 = add(x = var_509_cast_fp16, y = var_515_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_530_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171697664)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171700288)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor var_552_pad_type_0 = const()[name = tensor("op_552_pad_type_0"), val = tensor("valid")]; + tensor var_552_strides_0 = const()[name = tensor("op_552_strides_0"), val = tensor([1, 1])]; + tensor var_552_pad_0 = const()[name = tensor("op_552_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_552_dilations_0 = const()[name = tensor("op_552_dilations_0"), val = tensor([1, 1])]; + tensor var_552_groups_0 = const()[name = tensor("op_552_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(171702912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173341376))), name = tensor("layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173341952)))]; + tensor var_552_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_552_dilations_0, groups = var_552_groups_0, pad = var_552_pad_0, pad_type = var_552_pad_type_0, strides = var_552_strides_0, weight = layers_1_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_558_pad_type_0 = const()[name = tensor("op_558_pad_type_0"), val = tensor("valid")]; + tensor var_558_strides_0 = const()[name = tensor("op_558_strides_0"), val = tensor([1, 1])]; + tensor var_558_pad_0 = const()[name = tensor("op_558_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_558_dilations_0 = const()[name = tensor("op_558_dilations_0"), val = tensor([1, 1])]; + tensor var_558_groups_0 = const()[name = tensor("op_558_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173387648))), name = tensor("layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173344576))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_558_cast_fp16 = conv(dilations = var_558_dilations_0, groups = var_558_groups_0, pad = var_558_pad_0, pad_type = var_558_pad_type_0, strides = var_558_strides_0, weight = layers_1_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_23_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor query_7_cast_fp16 = add(x = var_552_cast_fp16, y = var_558_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_567_pad_type_0 = const()[name = tensor("op_567_pad_type_0"), val = tensor("valid")]; + tensor var_567_strides_0 = const()[name = tensor("op_567_strides_0"), val = tensor([1, 1])]; + tensor var_567_pad_0 = const()[name = tensor("op_567_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_567_dilations_0 = const()[name = tensor("op_567_dilations_0"), val = tensor([1, 1])]; + tensor var_567_groups_0 = const()[name = tensor("op_567_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173592512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175230976))), name = tensor("layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_567_cast_fp16 = conv(dilations = var_567_dilations_0, groups = var_567_groups_0, pad = var_567_pad_0, pad_type = var_567_pad_type_0, strides = var_567_strides_0, weight = layers_1_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_567_cast_fp16")]; + tensor var_573_pad_type_0 = const()[name = tensor("op_573_pad_type_0"), val = tensor("valid")]; + tensor var_573_strides_0 = const()[name = tensor("op_573_strides_0"), val = tensor([1, 1])]; + tensor var_573_pad_0 = const()[name = tensor("op_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_573_dilations_0 = const()[name = tensor("op_573_dilations_0"), val = tensor([1, 1])]; + tensor var_573_groups_0 = const()[name = tensor("op_573_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175274688))), name = tensor("layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175231552))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_573_cast_fp16 = conv(dilations = var_573_dilations_0, groups = var_573_groups_0, pad = var_573_pad_0, pad_type = var_573_pad_type_0, strides = var_573_strides_0, weight = layers_1_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_573_cast_fp16")]; + tensor key_7_cast_fp16 = add(x = var_567_cast_fp16, y = var_573_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor var_583_pad_type_0 = const()[name = tensor("op_583_pad_type_0"), val = tensor("valid")]; + tensor var_583_strides_0 = const()[name = tensor("op_583_strides_0"), val = tensor([1, 1])]; + tensor var_583_pad_0 = const()[name = tensor("op_583_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_583_dilations_0 = const()[name = tensor("op_583_dilations_0"), val = tensor([1, 1])]; + tensor var_583_groups_0 = const()[name = tensor("op_583_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175479552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177118016))), name = tensor("layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177118592)))]; + tensor var_583_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_583_dilations_0, groups = var_583_groups_0, pad = var_583_pad_0, pad_type = var_583_pad_type_0, strides = var_583_strides_0, weight = layers_1_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_583_cast_fp16")]; + tensor var_589_pad_type_0 = const()[name = tensor("op_589_pad_type_0"), val = tensor("valid")]; + tensor var_589_strides_0 = const()[name = tensor("op_589_strides_0"), val = tensor([1, 1])]; + tensor var_589_pad_0 = const()[name = tensor("op_589_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_589_dilations_0 = const()[name = tensor("op_589_dilations_0"), val = tensor([1, 1])]; + tensor var_589_groups_0 = const()[name = tensor("op_589_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177132352))), name = tensor("layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177121216))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_589_cast_fp16 = conv(dilations = var_589_dilations_0, groups = var_589_groups_0, pad = var_589_pad_0, pad_type = var_589_pad_type_0, strides = var_589_strides_0, weight = layers_1_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_589_cast_fp16")]; + tensor value_7_cast_fp16 = add(x = var_583_cast_fp16, y = var_589_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_593 = const()[name = tensor("op_593"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_593, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_595_to_fp16 = const()[name = tensor("op_595_to_fp16"), val = tensor(0x1p-3)]; + tensor var_596_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_595_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 20, 64, 1500])]; + tensor var_600_cast_fp16 = reshape(shape = var_599, x = key_7_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_596_cast_fp16, y = var_600_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_379, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 20, 64, 1500])]; + tensor var_605_cast_fp16 = reshape(shape = var_604, x = value_7_cast_fp16)[name = tensor("op_605_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_605_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 1280, 1, 1])]; + tensor input_13_cast_fp16 = reshape(shape = var_608, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_618_pad_type_0 = const()[name = tensor("op_618_pad_type_0"), val = tensor("valid")]; + tensor var_618_strides_0 = const()[name = tensor("op_618_strides_0"), val = tensor([1, 1])]; + tensor var_618_pad_0 = const()[name = tensor("op_618_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_618_dilations_0 = const()[name = tensor("op_618_dilations_0"), val = tensor([1, 1])]; + tensor var_618_groups_0 = const()[name = tensor("op_618_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177337216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178975680))), name = tensor("layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178976256)))]; + tensor var_618_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_618_dilations_0, groups = var_618_groups_0, pad = var_618_pad_0, pad_type = var_618_pad_type_0, strides = var_618_strides_0, weight = layers_1_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_624_pad_type_0 = const()[name = tensor("op_624_pad_type_0"), val = tensor("valid")]; + tensor var_624_strides_0 = const()[name = tensor("op_624_strides_0"), val = tensor([1, 1])]; + tensor var_624_pad_0 = const()[name = tensor("op_624_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_624_dilations_0 = const()[name = tensor("op_624_dilations_0"), val = tensor([1, 1])]; + tensor var_624_groups_0 = const()[name = tensor("op_624_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178989248))), name = tensor("layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178978880))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_624_cast_fp16 = conv(dilations = var_624_dilations_0, groups = var_624_groups_0, pad = var_624_pad_0, pad_type = var_624_pad_type_0, strides = var_624_strides_0, weight = layers_1_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_13_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor obj_25_cast_fp16 = add(x = var_618_cast_fp16, y = var_624_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_635_to_fp16 = const()[name = tensor("op_635_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_635_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179194112)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179196736)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor var_653_pad_type_0 = const()[name = tensor("op_653_pad_type_0"), val = tensor("valid")]; + tensor var_653_strides_0 = const()[name = tensor("op_653_strides_0"), val = tensor([1, 1])]; + tensor var_653_pad_0 = const()[name = tensor("op_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_653_dilations_0 = const()[name = tensor("op_653_dilations_0"), val = tensor([1, 1])]; + tensor var_653_groups_0 = const()[name = tensor("op_653_groups_0"), val = tensor(1)]; + tensor layers_1_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179199360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185753024))), name = tensor("layers_1_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_1_fc1_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_fc1_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185753600)))]; + tensor var_653_cast_fp16 = conv(bias = layers_1_fc1_inlier_module_bias_to_fp16, dilations = var_653_dilations_0, groups = var_653_groups_0, pad = var_653_pad_0, pad_type = var_653_pad_type_0, strides = var_653_strides_0, weight = layers_1_fc1_inlier_module_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor var_659_pad_type_0 = const()[name = tensor("op_659_pad_type_0"), val = tensor("valid")]; + tensor var_659_strides_0 = const()[name = tensor("op_659_strides_0"), val = tensor([1, 1])]; + tensor var_659_pad_0 = const()[name = tensor("op_659_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_659_dilations_0 = const()[name = tensor("op_659_dilations_0"), val = tensor([1, 1])]; + tensor var_659_groups_0 = const()[name = tensor("op_659_groups_0"), val = tensor(1)]; + tensor layers_1_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185848640))), name = tensor("layers_1_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185763904))), shape = tensor([5120, 1280, 1, 1])]; + tensor var_659_cast_fp16 = conv(dilations = var_659_dilations_0, groups = var_659_groups_0, pad = var_659_pad_0, pad_type = var_659_pad_type_0, strides = var_659_strides_0, weight = layers_1_fc1_outlier_module_weight_to_fp16_sparsified, x = input_15_cast_fp16)[name = tensor("op_659_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_653_cast_fp16, y = var_659_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_670_pad_type_0 = const()[name = tensor("op_670_pad_type_0"), val = tensor("valid")]; + tensor var_670_strides_0 = const()[name = tensor("op_670_strides_0"), val = tensor([1, 1])]; + tensor var_670_pad_0 = const()[name = tensor("op_670_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_670_dilations_0 = const()[name = tensor("op_670_dilations_0"), val = tensor([1, 1])]; + tensor var_670_groups_0 = const()[name = tensor("op_670_groups_0"), val = tensor(1)]; + tensor layers_1_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186667904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193221568))), name = tensor("layers_1_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_1_fc2_inlier_module_bias_to_fp16 = const()[name = tensor("layers_1_fc2_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193222144)))]; + tensor var_670_cast_fp16 = conv(bias = layers_1_fc2_inlier_module_bias_to_fp16, dilations = var_670_dilations_0, groups = var_670_groups_0, pad = var_670_pad_0, pad_type = var_670_pad_type_0, strides = var_670_strides_0, weight = layers_1_fc2_inlier_module_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor var_676_pad_type_0 = const()[name = tensor("op_676_pad_type_0"), val = tensor("valid")]; + tensor var_676_strides_0 = const()[name = tensor("op_676_strides_0"), val = tensor([1, 1])]; + tensor var_676_pad_0 = const()[name = tensor("op_676_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_676_dilations_0 = const()[name = tensor("op_676_dilations_0"), val = tensor([1, 1])]; + tensor var_676_groups_0 = const()[name = tensor("op_676_groups_0"), val = tensor(1)]; + tensor layers_1_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193312320))), name = tensor("layers_1_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193224768))), shape = tensor([1280, 5120, 1, 1])]; + tensor var_676_cast_fp16 = conv(dilations = var_676_dilations_0, groups = var_676_groups_0, pad = var_676_pad_0, pad_type = var_676_pad_type_0, strides = var_676_strides_0, weight = layers_1_fc2_outlier_module_weight_to_fp16_sparsified, x = input_19_cast_fp16)[name = tensor("op_676_cast_fp16")]; + tensor hidden_states_5_cast_fp16 = add(x = var_670_cast_fp16, y = var_676_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_713_to_fp16 = const()[name = tensor("op_713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_713_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194131584)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194134208)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor var_735_pad_type_0 = const()[name = tensor("op_735_pad_type_0"), val = tensor("valid")]; + tensor var_735_strides_0 = const()[name = tensor("op_735_strides_0"), val = tensor([1, 1])]; + tensor var_735_pad_0 = const()[name = tensor("op_735_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_735_dilations_0 = const()[name = tensor("op_735_dilations_0"), val = tensor([1, 1])]; + tensor var_735_groups_0 = const()[name = tensor("op_735_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194136832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195775296))), name = tensor("layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195775872)))]; + tensor var_735_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_735_dilations_0, groups = var_735_groups_0, pad = var_735_pad_0, pad_type = var_735_pad_type_0, strides = var_735_strides_0, weight = layers_2_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_741_pad_type_0 = const()[name = tensor("op_741_pad_type_0"), val = tensor("valid")]; + tensor var_741_strides_0 = const()[name = tensor("op_741_strides_0"), val = tensor([1, 1])]; + tensor var_741_pad_0 = const()[name = tensor("op_741_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_741_dilations_0 = const()[name = tensor("op_741_dilations_0"), val = tensor([1, 1])]; + tensor var_741_groups_0 = const()[name = tensor("op_741_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195811008))), name = tensor("layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195778496))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_741_cast_fp16 = conv(dilations = var_741_dilations_0, groups = var_741_groups_0, pad = var_741_pad_0, pad_type = var_741_pad_type_0, strides = var_741_strides_0, weight = layers_2_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor("op_741_cast_fp16")]; + tensor query_9_cast_fp16 = add(x = var_735_cast_fp16, y = var_741_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_750_pad_type_0 = const()[name = tensor("op_750_pad_type_0"), val = tensor("valid")]; + tensor var_750_strides_0 = const()[name = tensor("op_750_strides_0"), val = tensor([1, 1])]; + tensor var_750_pad_0 = const()[name = tensor("op_750_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_750_dilations_0 = const()[name = tensor("op_750_dilations_0"), val = tensor([1, 1])]; + tensor var_750_groups_0 = const()[name = tensor("op_750_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196015872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197654336))), name = tensor("layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_750_cast_fp16 = conv(dilations = var_750_dilations_0, groups = var_750_groups_0, pad = var_750_pad_0, pad_type = var_750_pad_type_0, strides = var_750_strides_0, weight = layers_2_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("op_750_cast_fp16")]; + tensor var_756_pad_type_0 = const()[name = tensor("op_756_pad_type_0"), val = tensor("valid")]; + tensor var_756_strides_0 = const()[name = tensor("op_756_strides_0"), val = tensor([1, 1])]; + tensor var_756_pad_0 = const()[name = tensor("op_756_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_756_dilations_0 = const()[name = tensor("op_756_dilations_0"), val = tensor([1, 1])]; + tensor var_756_groups_0 = const()[name = tensor("op_756_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197692288))), name = tensor("layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197654912))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_756_cast_fp16 = conv(dilations = var_756_dilations_0, groups = var_756_groups_0, pad = var_756_pad_0, pad_type = var_756_pad_type_0, strides = var_756_strides_0, weight = layers_2_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor("op_756_cast_fp16")]; + tensor current_key_5_cast_fp16 = add(x = var_750_cast_fp16, y = var_756_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_766_pad_type_0 = const()[name = tensor("op_766_pad_type_0"), val = tensor("valid")]; + tensor var_766_strides_0 = const()[name = tensor("op_766_strides_0"), val = tensor([1, 1])]; + tensor var_766_pad_0 = const()[name = tensor("op_766_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_766_dilations_0 = const()[name = tensor("op_766_dilations_0"), val = tensor([1, 1])]; + tensor var_766_groups_0 = const()[name = tensor("op_766_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197897152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199535616))), name = tensor("layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199536192)))]; + tensor var_766_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_766_dilations_0, groups = var_766_groups_0, pad = var_766_pad_0, pad_type = var_766_pad_type_0, strides = var_766_strides_0, weight = layers_2_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("op_766_cast_fp16")]; + tensor var_772_pad_type_0 = const()[name = tensor("op_772_pad_type_0"), val = tensor("valid")]; + tensor var_772_strides_0 = const()[name = tensor("op_772_strides_0"), val = tensor([1, 1])]; + tensor var_772_pad_0 = const()[name = tensor("op_772_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_772_dilations_0 = const()[name = tensor("op_772_dilations_0"), val = tensor([1, 1])]; + tensor var_772_groups_0 = const()[name = tensor("op_772_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199551744))), name = tensor("layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199538816))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_772_cast_fp16 = conv(dilations = var_772_dilations_0, groups = var_772_groups_0, pad = var_772_pad_0, pad_type = var_772_pad_type_0, strides = var_772_strides_0, weight = layers_2_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_29_cast_fp16)[name = tensor("op_772_cast_fp16")]; + tensor current_value_5_cast_fp16 = add(x = var_766_cast_fp16, y = var_772_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = var_53_cast_fp16_2, y = var_160_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor var_780_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_779_cast_fp16, y = var_780_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_783_cast_fp16 = mul(x = var_60_cast_fp16_2, y = var_160_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_784_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_784_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_783_cast_fp16, y = var_784_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_788 = const()[name = tensor("op_788"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_788, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_790_to_fp16 = const()[name = tensor("op_790_to_fp16"), val = tensor(0x1p-3)]; + tensor var_791_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_790_to_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_794 = const()[name = tensor("op_794"), val = tensor([1, 20, 64, 448])]; + tensor var_795_cast_fp16 = reshape(shape = var_794, x = key_9_cast_fp16)[name = tensor("op_795_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_791_cast_fp16, y = var_795_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_182_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_803_cast_fp16 = softmax(axis = var_688, x = mh_w_15_cast_fp16)[name = tensor("op_803_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 20, 64, 448])]; + tensor var_805_cast_fp16 = reshape(shape = var_804, x = value_9_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_805_cast_fp16, y = var_803_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_808 = const()[name = tensor("op_808"), val = tensor([1, 1280, 1, 1])]; + tensor input_21_cast_fp16 = reshape(shape = var_808, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_818_pad_type_0 = const()[name = tensor("op_818_pad_type_0"), val = tensor("valid")]; + tensor var_818_strides_0 = const()[name = tensor("op_818_strides_0"), val = tensor([1, 1])]; + tensor var_818_pad_0 = const()[name = tensor("op_818_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_818_dilations_0 = const()[name = tensor("op_818_dilations_0"), val = tensor([1, 1])]; + tensor var_818_groups_0 = const()[name = tensor("op_818_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199756608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201395072))), name = tensor("layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201395648)))]; + tensor var_818_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_818_dilations_0, groups = var_818_groups_0, pad = var_818_pad_0, pad_type = var_818_pad_type_0, strides = var_818_strides_0, weight = layers_2_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_824_pad_type_0 = const()[name = tensor("op_824_pad_type_0"), val = tensor("valid")]; + tensor var_824_strides_0 = const()[name = tensor("op_824_strides_0"), val = tensor([1, 1])]; + tensor var_824_pad_0 = const()[name = tensor("op_824_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_824_dilations_0 = const()[name = tensor("op_824_dilations_0"), val = tensor([1, 1])]; + tensor var_824_groups_0 = const()[name = tensor("op_824_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201409600))), name = tensor("layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201398272))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_824_cast_fp16 = conv(dilations = var_824_dilations_0, groups = var_824_groups_0, pad = var_824_pad_0, pad_type = var_824_pad_type_0, strides = var_824_strides_0, weight = layers_2_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_21_cast_fp16)[name = tensor("op_824_cast_fp16")]; + tensor obj_35_cast_fp16 = add(x = var_818_cast_fp16, y = var_824_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_839_to_fp16 = const()[name = tensor("op_839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_839_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201614464)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201617088)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor var_861_pad_type_0 = const()[name = tensor("op_861_pad_type_0"), val = tensor("valid")]; + tensor var_861_strides_0 = const()[name = tensor("op_861_strides_0"), val = tensor([1, 1])]; + tensor var_861_pad_0 = const()[name = tensor("op_861_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_861_dilations_0 = const()[name = tensor("op_861_dilations_0"), val = tensor([1, 1])]; + tensor var_861_groups_0 = const()[name = tensor("op_861_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201619712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203258176))), name = tensor("layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203258752)))]; + tensor var_861_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_861_dilations_0, groups = var_861_groups_0, pad = var_861_pad_0, pad_type = var_861_pad_type_0, strides = var_861_strides_0, weight = layers_2_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_867_pad_type_0 = const()[name = tensor("op_867_pad_type_0"), val = tensor("valid")]; + tensor var_867_strides_0 = const()[name = tensor("op_867_strides_0"), val = tensor([1, 1])]; + tensor var_867_pad_0 = const()[name = tensor("op_867_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_867_dilations_0 = const()[name = tensor("op_867_dilations_0"), val = tensor([1, 1])]; + tensor var_867_groups_0 = const()[name = tensor("op_867_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203289600))), name = tensor("layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203261376))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_867_cast_fp16 = conv(dilations = var_867_dilations_0, groups = var_867_groups_0, pad = var_867_pad_0, pad_type = var_867_pad_type_0, strides = var_867_strides_0, weight = layers_2_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_37_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor query_11_cast_fp16 = add(x = var_861_cast_fp16, y = var_867_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_876_pad_type_0 = const()[name = tensor("op_876_pad_type_0"), val = tensor("valid")]; + tensor var_876_strides_0 = const()[name = tensor("op_876_strides_0"), val = tensor([1, 1])]; + tensor var_876_pad_0 = const()[name = tensor("op_876_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_876_dilations_0 = const()[name = tensor("op_876_dilations_0"), val = tensor([1, 1])]; + tensor var_876_groups_0 = const()[name = tensor("op_876_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(203494464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205132928))), name = tensor("layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_876_cast_fp16 = conv(dilations = var_876_dilations_0, groups = var_876_groups_0, pad = var_876_pad_0, pad_type = var_876_pad_type_0, strides = var_876_strides_0, weight = layers_2_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_876_cast_fp16")]; + tensor var_882_pad_type_0 = const()[name = tensor("op_882_pad_type_0"), val = tensor("valid")]; + tensor var_882_strides_0 = const()[name = tensor("op_882_strides_0"), val = tensor([1, 1])]; + tensor var_882_pad_0 = const()[name = tensor("op_882_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_882_dilations_0 = const()[name = tensor("op_882_dilations_0"), val = tensor([1, 1])]; + tensor var_882_groups_0 = const()[name = tensor("op_882_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205161152))), name = tensor("layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205133504))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_882_cast_fp16 = conv(dilations = var_882_dilations_0, groups = var_882_groups_0, pad = var_882_pad_0, pad_type = var_882_pad_type_0, strides = var_882_strides_0, weight = layers_2_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_882_cast_fp16")]; + tensor key_11_cast_fp16 = add(x = var_876_cast_fp16, y = var_882_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor var_892_pad_type_0 = const()[name = tensor("op_892_pad_type_0"), val = tensor("valid")]; + tensor var_892_strides_0 = const()[name = tensor("op_892_strides_0"), val = tensor([1, 1])]; + tensor var_892_pad_0 = const()[name = tensor("op_892_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_892_dilations_0 = const()[name = tensor("op_892_dilations_0"), val = tensor([1, 1])]; + tensor var_892_groups_0 = const()[name = tensor("op_892_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205366016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207004480))), name = tensor("layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207005056)))]; + tensor var_892_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_892_dilations_0, groups = var_892_groups_0, pad = var_892_pad_0, pad_type = var_892_pad_type_0, strides = var_892_strides_0, weight = layers_2_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_892_cast_fp16")]; + tensor var_898_pad_type_0 = const()[name = tensor("op_898_pad_type_0"), val = tensor("valid")]; + tensor var_898_strides_0 = const()[name = tensor("op_898_strides_0"), val = tensor([1, 1])]; + tensor var_898_pad_0 = const()[name = tensor("op_898_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_898_dilations_0 = const()[name = tensor("op_898_dilations_0"), val = tensor([1, 1])]; + tensor var_898_groups_0 = const()[name = tensor("op_898_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207019264))), name = tensor("layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207007680))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_898_cast_fp16 = conv(dilations = var_898_dilations_0, groups = var_898_groups_0, pad = var_898_pad_0, pad_type = var_898_pad_type_0, strides = var_898_strides_0, weight = layers_2_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_898_cast_fp16")]; + tensor value_11_cast_fp16 = add(x = var_892_cast_fp16, y = var_898_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_902 = const()[name = tensor("op_902"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_902, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_904_to_fp16 = const()[name = tensor("op_904_to_fp16"), val = tensor(0x1p-3)]; + tensor var_905_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_904_to_fp16)[name = tensor("op_905_cast_fp16")]; + tensor var_908 = const()[name = tensor("op_908"), val = tensor([1, 20, 64, 1500])]; + tensor var_909_cast_fp16 = reshape(shape = var_908, x = key_11_cast_fp16)[name = tensor("op_909_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_905_cast_fp16, y = var_909_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_688, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_913 = const()[name = tensor("op_913"), val = tensor([1, 20, 64, 1500])]; + tensor var_914_cast_fp16 = reshape(shape = var_913, x = value_11_cast_fp16)[name = tensor("op_914_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_914_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_917 = const()[name = tensor("op_917"), val = tensor([1, 1280, 1, 1])]; + tensor input_23_cast_fp16 = reshape(shape = var_917, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_927_pad_type_0 = const()[name = tensor("op_927_pad_type_0"), val = tensor("valid")]; + tensor var_927_strides_0 = const()[name = tensor("op_927_strides_0"), val = tensor([1, 1])]; + tensor var_927_pad_0 = const()[name = tensor("op_927_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_927_dilations_0 = const()[name = tensor("op_927_dilations_0"), val = tensor([1, 1])]; + tensor var_927_groups_0 = const()[name = tensor("op_927_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207224128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208862592))), name = tensor("layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208863168)))]; + tensor var_927_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_927_dilations_0, groups = var_927_groups_0, pad = var_927_pad_0, pad_type = var_927_pad_type_0, strides = var_927_strides_0, weight = layers_2_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("op_927_cast_fp16")]; + tensor var_933_pad_type_0 = const()[name = tensor("op_933_pad_type_0"), val = tensor("valid")]; + tensor var_933_strides_0 = const()[name = tensor("op_933_strides_0"), val = tensor([1, 1])]; + tensor var_933_pad_0 = const()[name = tensor("op_933_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_933_dilations_0 = const()[name = tensor("op_933_dilations_0"), val = tensor([1, 1])]; + tensor var_933_groups_0 = const()[name = tensor("op_933_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208878592))), name = tensor("layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208865792))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_933_cast_fp16 = conv(dilations = var_933_dilations_0, groups = var_933_groups_0, pad = var_933_pad_0, pad_type = var_933_pad_type_0, strides = var_933_strides_0, weight = layers_2_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_23_cast_fp16)[name = tensor("op_933_cast_fp16")]; + tensor obj_39_cast_fp16 = add(x = var_927_cast_fp16, y = var_933_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_947_to_fp16 = const()[name = tensor("op_947_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_947_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209083456)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209086080)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor var_965_pad_type_0 = const()[name = tensor("op_965_pad_type_0"), val = tensor("valid")]; + tensor var_965_strides_0 = const()[name = tensor("op_965_strides_0"), val = tensor([1, 1])]; + tensor var_965_pad_0 = const()[name = tensor("op_965_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_965_dilations_0 = const()[name = tensor("op_965_dilations_0"), val = tensor([1, 1])]; + tensor var_965_groups_0 = const()[name = tensor("op_965_groups_0"), val = tensor(1)]; + tensor layers_2_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209088704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215642368))), name = tensor("layers_2_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_2_fc1_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_fc1_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215642944)))]; + tensor var_965_cast_fp16 = conv(bias = layers_2_fc1_inlier_module_bias_to_fp16, dilations = var_965_dilations_0, groups = var_965_groups_0, pad = var_965_pad_0, pad_type = var_965_pad_type_0, strides = var_965_strides_0, weight = layers_2_fc1_inlier_module_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor var_971_pad_type_0 = const()[name = tensor("op_971_pad_type_0"), val = tensor("valid")]; + tensor var_971_strides_0 = const()[name = tensor("op_971_strides_0"), val = tensor([1, 1])]; + tensor var_971_pad_0 = const()[name = tensor("op_971_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_971_dilations_0 = const()[name = tensor("op_971_dilations_0"), val = tensor([1, 1])]; + tensor var_971_groups_0 = const()[name = tensor("op_971_groups_0"), val = tensor(1)]; + tensor layers_2_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215812672))), name = tensor("layers_2_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215653248))), shape = tensor([5120, 1280, 1, 1])]; + tensor var_971_cast_fp16 = conv(dilations = var_971_dilations_0, groups = var_971_groups_0, pad = var_971_pad_0, pad_type = var_971_pad_type_0, strides = var_971_strides_0, weight = layers_2_fc1_outlier_module_weight_to_fp16_sparsified, x = input_25_cast_fp16)[name = tensor("op_971_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = var_965_cast_fp16, y = var_971_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_982_pad_type_0 = const()[name = tensor("op_982_pad_type_0"), val = tensor("valid")]; + tensor var_982_strides_0 = const()[name = tensor("op_982_strides_0"), val = tensor([1, 1])]; + tensor var_982_pad_0 = const()[name = tensor("op_982_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_982_dilations_0 = const()[name = tensor("op_982_dilations_0"), val = tensor([1, 1])]; + tensor var_982_groups_0 = const()[name = tensor("op_982_groups_0"), val = tensor(1)]; + tensor layers_2_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216631936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223185600))), name = tensor("layers_2_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_2_fc2_inlier_module_bias_to_fp16 = const()[name = tensor("layers_2_fc2_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223186176)))]; + tensor var_982_cast_fp16 = conv(bias = layers_2_fc2_inlier_module_bias_to_fp16, dilations = var_982_dilations_0, groups = var_982_groups_0, pad = var_982_pad_0, pad_type = var_982_pad_type_0, strides = var_982_strides_0, weight = layers_2_fc2_inlier_module_weight_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("op_982_cast_fp16")]; + tensor var_988_pad_type_0 = const()[name = tensor("op_988_pad_type_0"), val = tensor("valid")]; + tensor var_988_strides_0 = const()[name = tensor("op_988_strides_0"), val = tensor([1, 1])]; + tensor var_988_pad_0 = const()[name = tensor("op_988_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_988_dilations_0 = const()[name = tensor("op_988_dilations_0"), val = tensor([1, 1])]; + tensor var_988_groups_0 = const()[name = tensor("op_988_groups_0"), val = tensor(1)]; + tensor layers_2_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223266112))), name = tensor("layers_2_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223188800))), shape = tensor([1280, 5120, 1, 1])]; + tensor var_988_cast_fp16 = conv(dilations = var_988_dilations_0, groups = var_988_groups_0, pad = var_988_pad_0, pad_type = var_988_pad_type_0, strides = var_988_strides_0, weight = layers_2_fc2_outlier_module_weight_to_fp16_sparsified, x = input_29_cast_fp16)[name = tensor("op_988_cast_fp16")]; + tensor hidden_states_7_cast_fp16 = add(x = var_982_cast_fp16, y = var_988_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_1026_to_fp16 = const()[name = tensor("op_1026_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1026_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224085376)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224088000)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor var_1048_pad_type_0 = const()[name = tensor("op_1048_pad_type_0"), val = tensor("valid")]; + tensor var_1048_strides_0 = const()[name = tensor("op_1048_strides_0"), val = tensor([1, 1])]; + tensor var_1048_pad_0 = const()[name = tensor("op_1048_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1048_dilations_0 = const()[name = tensor("op_1048_dilations_0"), val = tensor([1, 1])]; + tensor var_1048_groups_0 = const()[name = tensor("op_1048_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224090624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225729088))), name = tensor("layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225729664)))]; + tensor var_1048_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1048_dilations_0, groups = var_1048_groups_0, pad = var_1048_pad_0, pad_type = var_1048_pad_type_0, strides = var_1048_strides_0, weight = layers_3_self_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("op_1048_cast_fp16")]; + tensor var_1054_pad_type_0 = const()[name = tensor("op_1054_pad_type_0"), val = tensor("valid")]; + tensor var_1054_strides_0 = const()[name = tensor("op_1054_strides_0"), val = tensor([1, 1])]; + tensor var_1054_pad_0 = const()[name = tensor("op_1054_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1054_dilations_0 = const()[name = tensor("op_1054_dilations_0"), val = tensor([1, 1])]; + tensor var_1054_groups_0 = const()[name = tensor("op_1054_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225753728))), name = tensor("layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225732288))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1054_cast_fp16 = conv(dilations = var_1054_dilations_0, groups = var_1054_groups_0, pad = var_1054_pad_0, pad_type = var_1054_pad_type_0, strides = var_1054_strides_0, weight = layers_3_self_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor query_13_cast_fp16 = add(x = var_1048_cast_fp16, y = var_1054_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_1063_pad_type_0 = const()[name = tensor("op_1063_pad_type_0"), val = tensor("valid")]; + tensor var_1063_strides_0 = const()[name = tensor("op_1063_strides_0"), val = tensor([1, 1])]; + tensor var_1063_pad_0 = const()[name = tensor("op_1063_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1063_dilations_0 = const()[name = tensor("op_1063_dilations_0"), val = tensor([1, 1])]; + tensor var_1063_groups_0 = const()[name = tensor("op_1063_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225958592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227597056))), name = tensor("layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1063_cast_fp16 = conv(dilations = var_1063_dilations_0, groups = var_1063_groups_0, pad = var_1063_pad_0, pad_type = var_1063_pad_type_0, strides = var_1063_strides_0, weight = layers_3_self_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor var_1069_pad_type_0 = const()[name = tensor("op_1069_pad_type_0"), val = tensor("valid")]; + tensor var_1069_strides_0 = const()[name = tensor("op_1069_strides_0"), val = tensor([1, 1])]; + tensor var_1069_pad_0 = const()[name = tensor("op_1069_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1069_dilations_0 = const()[name = tensor("op_1069_dilations_0"), val = tensor([1, 1])]; + tensor var_1069_groups_0 = const()[name = tensor("op_1069_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227618368))), name = tensor("layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227597632))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1069_cast_fp16 = conv(dilations = var_1069_dilations_0, groups = var_1069_groups_0, pad = var_1069_pad_0, pad_type = var_1069_pad_type_0, strides = var_1069_strides_0, weight = layers_3_self_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor("op_1069_cast_fp16")]; + tensor current_key_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1069_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_1079_pad_type_0 = const()[name = tensor("op_1079_pad_type_0"), val = tensor("valid")]; + tensor var_1079_strides_0 = const()[name = tensor("op_1079_strides_0"), val = tensor([1, 1])]; + tensor var_1079_pad_0 = const()[name = tensor("op_1079_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1079_dilations_0 = const()[name = tensor("op_1079_dilations_0"), val = tensor([1, 1])]; + tensor var_1079_groups_0 = const()[name = tensor("op_1079_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227823232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229461696))), name = tensor("layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229462272)))]; + tensor var_1079_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1079_dilations_0, groups = var_1079_groups_0, pad = var_1079_pad_0, pad_type = var_1079_pad_type_0, strides = var_1079_strides_0, weight = layers_3_self_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("op_1079_cast_fp16")]; + tensor var_1085_pad_type_0 = const()[name = tensor("op_1085_pad_type_0"), val = tensor("valid")]; + tensor var_1085_strides_0 = const()[name = tensor("op_1085_strides_0"), val = tensor([1, 1])]; + tensor var_1085_pad_0 = const()[name = tensor("op_1085_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1085_dilations_0 = const()[name = tensor("op_1085_dilations_0"), val = tensor([1, 1])]; + tensor var_1085_groups_0 = const()[name = tensor("op_1085_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229479424))), name = tensor("layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229464896))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1085_cast_fp16 = conv(dilations = var_1085_dilations_0, groups = var_1085_groups_0, pad = var_1085_pad_0, pad_type = var_1085_pad_type_0, strides = var_1085_strides_0, weight = layers_3_self_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = obj_43_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor current_value_cast_fp16 = add(x = var_1079_cast_fp16, y = var_1085_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_53_cast_fp16_3, y = var_160_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = current_key_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_60_cast_fp16_3, y = var_160_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = current_value_cast_fp16, y = var_158_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_1101, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_1103_to_fp16 = const()[name = tensor("op_1103_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1104_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1103_to_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 20, 64, 448])]; + tensor var_1108_cast_fp16 = reshape(shape = var_1107, x = key_13_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1104_cast_fp16, y = var_1108_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_182_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_1116_cast_fp16 = softmax(axis = var_1001, x = mh_w_21_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([1, 20, 64, 448])]; + tensor var_1118_cast_fp16 = reshape(shape = var_1117, x = value_13_cast_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1118_cast_fp16, y = var_1116_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1121 = const()[name = tensor("op_1121"), val = tensor([1, 1280, 1, 1])]; + tensor input_31_cast_fp16 = reshape(shape = var_1121, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_1131_pad_type_0 = const()[name = tensor("op_1131_pad_type_0"), val = tensor("valid")]; + tensor var_1131_strides_0 = const()[name = tensor("op_1131_strides_0"), val = tensor([1, 1])]; + tensor var_1131_pad_0 = const()[name = tensor("op_1131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1131_dilations_0 = const()[name = tensor("op_1131_dilations_0"), val = tensor([1, 1])]; + tensor var_1131_groups_0 = const()[name = tensor("op_1131_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229684288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231322752))), name = tensor("layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231323328)))]; + tensor var_1131_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1131_dilations_0, groups = var_1131_groups_0, pad = var_1131_pad_0, pad_type = var_1131_pad_type_0, strides = var_1131_strides_0, weight = layers_3_self_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor("op_1131_cast_fp16")]; + tensor var_1137_pad_type_0 = const()[name = tensor("op_1137_pad_type_0"), val = tensor("valid")]; + tensor var_1137_strides_0 = const()[name = tensor("op_1137_strides_0"), val = tensor([1, 1])]; + tensor var_1137_pad_0 = const()[name = tensor("op_1137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1137_dilations_0 = const()[name = tensor("op_1137_dilations_0"), val = tensor([1, 1])]; + tensor var_1137_groups_0 = const()[name = tensor("op_1137_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231340224))), name = tensor("layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231325952))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1137_cast_fp16 = conv(dilations = var_1137_dilations_0, groups = var_1137_groups_0, pad = var_1137_pad_0, pad_type = var_1137_pad_type_0, strides = var_1137_strides_0, weight = layers_3_self_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_31_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor obj_49_cast_fp16 = add(x = var_1131_cast_fp16, y = var_1137_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1152_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231545088)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231547712)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor var_1174_pad_type_0 = const()[name = tensor("op_1174_pad_type_0"), val = tensor("valid")]; + tensor var_1174_strides_0 = const()[name = tensor("op_1174_strides_0"), val = tensor([1, 1])]; + tensor var_1174_pad_0 = const()[name = tensor("op_1174_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1174_dilations_0 = const()[name = tensor("op_1174_dilations_0"), val = tensor([1, 1])]; + tensor var_1174_groups_0 = const()[name = tensor("op_1174_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231550336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233188800))), name = tensor("layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233189376)))]; + tensor var_1174_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_inlier_module_bias_to_fp16, dilations = var_1174_dilations_0, groups = var_1174_groups_0, pad = var_1174_pad_0, pad_type = var_1174_pad_type_0, strides = var_1174_strides_0, weight = layers_3_encoder_attn_q_proj_inlier_module_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1180_pad_type_0 = const()[name = tensor("op_1180_pad_type_0"), val = tensor("valid")]; + tensor var_1180_strides_0 = const()[name = tensor("op_1180_strides_0"), val = tensor([1, 1])]; + tensor var_1180_pad_0 = const()[name = tensor("op_1180_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1180_dilations_0 = const()[name = tensor("op_1180_dilations_0"), val = tensor([1, 1])]; + tensor var_1180_groups_0 = const()[name = tensor("op_1180_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233207552))), name = tensor("layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233192000))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1180_cast_fp16 = conv(dilations = var_1180_dilations_0, groups = var_1180_groups_0, pad = var_1180_pad_0, pad_type = var_1180_pad_type_0, strides = var_1180_strides_0, weight = layers_3_encoder_attn_q_proj_outlier_module_weight_to_fp16_sparsified, x = obj_51_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor query_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1180_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_1189_pad_type_0 = const()[name = tensor("op_1189_pad_type_0"), val = tensor("valid")]; + tensor var_1189_strides_0 = const()[name = tensor("op_1189_strides_0"), val = tensor([1, 1])]; + tensor var_1189_pad_0 = const()[name = tensor("op_1189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1189_dilations_0 = const()[name = tensor("op_1189_dilations_0"), val = tensor([1, 1])]; + tensor var_1189_groups_0 = const()[name = tensor("op_1189_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233412416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235050880))), name = tensor("layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1189_cast_fp16 = conv(dilations = var_1189_dilations_0, groups = var_1189_groups_0, pad = var_1189_pad_0, pad_type = var_1189_pad_type_0, strides = var_1189_strides_0, weight = layers_3_encoder_attn_k_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_1189_cast_fp16")]; + tensor var_1195_pad_type_0 = const()[name = tensor("op_1195_pad_type_0"), val = tensor("valid")]; + tensor var_1195_strides_0 = const()[name = tensor("op_1195_strides_0"), val = tensor([1, 1])]; + tensor var_1195_pad_0 = const()[name = tensor("op_1195_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1195_dilations_0 = const()[name = tensor("op_1195_dilations_0"), val = tensor([1, 1])]; + tensor var_1195_groups_0 = const()[name = tensor("op_1195_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235074048))), name = tensor("layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051456))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1195_cast_fp16 = conv(dilations = var_1195_dilations_0, groups = var_1195_groups_0, pad = var_1195_pad_0, pad_type = var_1195_pad_type_0, strides = var_1195_strides_0, weight = layers_3_encoder_attn_k_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_1195_cast_fp16")]; + tensor key_cast_fp16 = add(x = var_1189_cast_fp16, y = var_1195_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor var_1205_pad_type_0 = const()[name = tensor("op_1205_pad_type_0"), val = tensor("valid")]; + tensor var_1205_strides_0 = const()[name = tensor("op_1205_strides_0"), val = tensor([1, 1])]; + tensor var_1205_pad_0 = const()[name = tensor("op_1205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1205_dilations_0 = const()[name = tensor("op_1205_dilations_0"), val = tensor([1, 1])]; + tensor var_1205_groups_0 = const()[name = tensor("op_1205_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235278912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236917376))), name = tensor("layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236917952)))]; + tensor var_1205_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_inlier_module_bias_to_fp16, dilations = var_1205_dilations_0, groups = var_1205_groups_0, pad = var_1205_pad_0, pad_type = var_1205_pad_type_0, strides = var_1205_strides_0, weight = layers_3_encoder_attn_v_proj_inlier_module_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_1205_cast_fp16")]; + tensor var_1211_pad_type_0 = const()[name = tensor("op_1211_pad_type_0"), val = tensor("valid")]; + tensor var_1211_strides_0 = const()[name = tensor("op_1211_strides_0"), val = tensor([1, 1])]; + tensor var_1211_pad_0 = const()[name = tensor("op_1211_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1211_dilations_0 = const()[name = tensor("op_1211_dilations_0"), val = tensor([1, 1])]; + tensor var_1211_groups_0 = const()[name = tensor("op_1211_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236934272))), name = tensor("layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236920576))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1211_cast_fp16 = conv(dilations = var_1211_dilations_0, groups = var_1211_groups_0, pad = var_1211_pad_0, pad_type = var_1211_pad_type_0, strides = var_1211_strides_0, weight = layers_3_encoder_attn_v_proj_outlier_module_weight_to_fp16_sparsified, x = encoder_output_embeds)[name = tensor("op_1211_cast_fp16")]; + tensor value_cast_fp16 = add(x = var_1205_cast_fp16, y = var_1211_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([1, 20, 64, 1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_1215, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_1217_to_fp16 = const()[name = tensor("op_1217_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1218_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1217_to_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 20, 64, 1500])]; + tensor var_1222_cast_fp16 = reshape(shape = var_1221, x = key_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1218_cast_fp16, y = var_1222_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_1001, x = mh_w_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_1226 = const()[name = tensor("op_1226"), val = tensor([1, 20, 64, 1500])]; + tensor var_1227_cast_fp16 = reshape(shape = var_1226, x = value_cast_fp16)[name = tensor("op_1227_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1227_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1280, 1, 1])]; + tensor input_33_cast_fp16 = reshape(shape = var_1230, x = attn_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_1240_pad_type_0 = const()[name = tensor("op_1240_pad_type_0"), val = tensor("valid")]; + tensor var_1240_strides_0 = const()[name = tensor("op_1240_strides_0"), val = tensor([1, 1])]; + tensor var_1240_pad_0 = const()[name = tensor("op_1240_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1240_dilations_0 = const()[name = tensor("op_1240_dilations_0"), val = tensor([1, 1])]; + tensor var_1240_groups_0 = const()[name = tensor("op_1240_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237139136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238777600))), name = tensor("layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238778176)))]; + tensor var_1240_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_inlier_module_bias_to_fp16, dilations = var_1240_dilations_0, groups = var_1240_groups_0, pad = var_1240_pad_0, pad_type = var_1240_pad_type_0, strides = var_1240_strides_0, weight = layers_3_encoder_attn_o_proj_inlier_module_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1246_pad_type_0 = const()[name = tensor("op_1246_pad_type_0"), val = tensor("valid")]; + tensor var_1246_strides_0 = const()[name = tensor("op_1246_strides_0"), val = tensor([1, 1])]; + tensor var_1246_pad_0 = const()[name = tensor("op_1246_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1246_dilations_0 = const()[name = tensor("op_1246_dilations_0"), val = tensor([1, 1])]; + tensor var_1246_groups_0 = const()[name = tensor("op_1246_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238792320))), name = tensor("layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238780800))), shape = tensor([1280, 1280, 1, 1])]; + tensor var_1246_cast_fp16 = conv(dilations = var_1246_dilations_0, groups = var_1246_groups_0, pad = var_1246_pad_0, pad_type = var_1246_pad_type_0, strides = var_1246_strides_0, weight = layers_3_encoder_attn_o_proj_outlier_module_weight_to_fp16_sparsified, x = input_33_cast_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor obj_53_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1246_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_1260_to_fp16 = const()[name = tensor("op_1260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1260_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238997184)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238999808)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor var_1278_pad_type_0 = const()[name = tensor("op_1278_pad_type_0"), val = tensor("valid")]; + tensor var_1278_strides_0 = const()[name = tensor("op_1278_strides_0"), val = tensor([1, 1])]; + tensor var_1278_pad_0 = const()[name = tensor("op_1278_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1278_dilations_0 = const()[name = tensor("op_1278_dilations_0"), val = tensor([1, 1])]; + tensor var_1278_groups_0 = const()[name = tensor("op_1278_groups_0"), val = tensor(1)]; + tensor layers_3_fc1_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239002432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245556096))), name = tensor("layers_3_fc1_inlier_module_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_3_fc1_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_fc1_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245556672)))]; + tensor var_1278_cast_fp16 = conv(bias = layers_3_fc1_inlier_module_bias_to_fp16, dilations = var_1278_dilations_0, groups = var_1278_groups_0, pad = var_1278_pad_0, pad_type = var_1278_pad_type_0, strides = var_1278_strides_0, weight = layers_3_fc1_inlier_module_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1284_pad_type_0 = const()[name = tensor("op_1284_pad_type_0"), val = tensor("valid")]; + tensor var_1284_strides_0 = const()[name = tensor("op_1284_strides_0"), val = tensor([1, 1])]; + tensor var_1284_pad_0 = const()[name = tensor("op_1284_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1284_dilations_0 = const()[name = tensor("op_1284_dilations_0"), val = tensor([1, 1])]; + tensor var_1284_groups_0 = const()[name = tensor("op_1284_groups_0"), val = tensor(1)]; + tensor layers_3_fc1_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245620096))), name = tensor("layers_3_fc1_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245566976))), shape = tensor([5120, 1280, 1, 1])]; + tensor var_1284_cast_fp16 = conv(dilations = var_1284_dilations_0, groups = var_1284_groups_0, pad = var_1284_pad_0, pad_type = var_1284_pad_type_0, strides = var_1284_strides_0, weight = layers_3_fc1_outlier_module_weight_to_fp16_sparsified, x = input_35_cast_fp16)[name = tensor("op_1284_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1284_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_1295_pad_type_0 = const()[name = tensor("op_1295_pad_type_0"), val = tensor("valid")]; + tensor var_1295_strides_0 = const()[name = tensor("op_1295_strides_0"), val = tensor([1, 1])]; + tensor var_1295_pad_0 = const()[name = tensor("op_1295_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1295_dilations_0 = const()[name = tensor("op_1295_dilations_0"), val = tensor([1, 1])]; + tensor var_1295_groups_0 = const()[name = tensor("op_1295_groups_0"), val = tensor(1)]; + tensor layers_3_fc2_inlier_module_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246439360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252993024))), name = tensor("layers_3_fc2_inlier_module_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_3_fc2_inlier_module_bias_to_fp16 = const()[name = tensor("layers_3_fc2_inlier_module_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252993600)))]; + tensor var_1295_cast_fp16 = conv(bias = layers_3_fc2_inlier_module_bias_to_fp16, dilations = var_1295_dilations_0, groups = var_1295_groups_0, pad = var_1295_pad_0, pad_type = var_1295_pad_type_0, strides = var_1295_strides_0, weight = layers_3_fc2_inlier_module_weight_to_fp16_palettized, x = input_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1301_pad_type_0 = const()[name = tensor("op_1301_pad_type_0"), val = tensor("valid")]; + tensor var_1301_strides_0 = const()[name = tensor("op_1301_strides_0"), val = tensor([1, 1])]; + tensor var_1301_pad_0 = const()[name = tensor("op_1301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1301_dilations_0 = const()[name = tensor("op_1301_dilations_0"), val = tensor([1, 1])]; + tensor var_1301_groups_0 = const()[name = tensor("op_1301_groups_0"), val = tensor(1)]; + tensor layers_3_fc2_outlier_module_weight_to_fp16_sparsified = constexpr_sparse_to_dense()[mask = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253066816))), name = tensor("layers_3_fc2_outlier_module_weight_to_fp16_sparsified"), nonzero_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252996224))), shape = tensor([1280, 5120, 1, 1])]; + tensor var_1301_cast_fp16 = conv(dilations = var_1301_dilations_0, groups = var_1301_groups_0, pad = var_1301_pad_0, pad_type = var_1301_pad_type_0, strides = var_1301_strides_0, weight = layers_3_fc2_outlier_module_weight_to_fp16_sparsified, x = input_cast_fp16)[name = tensor("op_1301_cast_fp16")]; + tensor hidden_states_9_cast_fp16 = add(x = var_1295_cast_fp16, y = var_1301_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_1321_to_fp16 = const()[name = tensor("op_1321_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1321_to_fp16, x = inputs_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253886080)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253888704)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_1332_axes_0 = const()[name = tensor("op_1332_axes_0"), val = tensor([2])]; + tensor var_1332_cast_fp16 = squeeze(axes = var_1332_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253891328)))]; + tensor var_1335_cast_fp16 = transpose(perm = var_1335_perm_0, x = var_1332_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1335_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_1339 = const()[name = tensor("op_1339"), val = tensor(1)]; + tensor obj_59_interleave_0 = const()[name = tensor("obj_59_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_1339, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor("obj_59_cast_fp16")]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor(1)]; + tensor obj_61_interleave_0 = const()[name = tensor("obj_61_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_1342, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor("obj_61_cast_fp16")]; + tensor var_1353_begin_0 = const()[name = tensor("op_1353_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_1353_end_0 = const()[name = tensor("op_1353_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_1353_end_mask_0 = const()[name = tensor("op_1353_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1353_cast_fp16 = slice_by_index(begin = var_1353_begin_0, end = var_1353_end_0, end_mask = var_1353_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1353_cast_fp16")]; + tensor var_1356_begin_0 = const()[name = tensor("op_1356_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1356_end_0 = const()[name = tensor("op_1356_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1356_end_mask_0 = const()[name = tensor("op_1356_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1356_squeeze_mask_0 = const()[name = tensor("op_1356_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1356_cast_fp16 = slice_by_index(begin = var_1356_begin_0, end = var_1356_end_0, end_mask = var_1356_end_mask_0, squeeze_mask = var_1356_squeeze_mask_0, x = var_1353_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor var_1371_begin_0 = const()[name = tensor("op_1371_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1371_end_0 = const()[name = tensor("op_1371_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_1371_end_mask_0 = const()[name = tensor("op_1371_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1371_cast_fp16 = slice_by_index(begin = var_1371_begin_0, end = var_1371_end_0, end_mask = var_1371_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1374_begin_0 = const()[name = tensor("op_1374_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1374_end_0 = const()[name = tensor("op_1374_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1374_end_mask_0 = const()[name = tensor("op_1374_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1374_squeeze_mask_0 = const()[name = tensor("op_1374_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1374_cast_fp16 = slice_by_index(begin = var_1374_begin_0, end = var_1374_end_0, end_mask = var_1374_end_mask_0, squeeze_mask = var_1374_squeeze_mask_0, x = var_1371_cast_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor var_1389_begin_0 = const()[name = tensor("op_1389_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_1389_end_0 = const()[name = tensor("op_1389_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_1389_end_mask_0 = const()[name = tensor("op_1389_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1389_cast_fp16 = slice_by_index(begin = var_1389_begin_0, end = var_1389_end_0, end_mask = var_1389_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor var_1392_begin_0 = const()[name = tensor("op_1392_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1392_end_0 = const()[name = tensor("op_1392_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1392_end_mask_0 = const()[name = tensor("op_1392_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1392_squeeze_mask_0 = const()[name = tensor("op_1392_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1392_cast_fp16 = slice_by_index(begin = var_1392_begin_0, end = var_1392_end_0, end_mask = var_1392_end_mask_0, squeeze_mask = var_1392_squeeze_mask_0, x = var_1389_cast_fp16)[name = tensor("op_1392_cast_fp16")]; + tensor var_1407_begin_0 = const()[name = tensor("op_1407_begin_0"), val = tensor([0, 6, 0, 0])]; + tensor var_1407_end_0 = const()[name = tensor("op_1407_end_0"), val = tensor([1, 7, 1, 1500])]; + tensor var_1407_end_mask_0 = const()[name = tensor("op_1407_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1407_cast_fp16 = slice_by_index(begin = var_1407_begin_0, end = var_1407_end_0, end_mask = var_1407_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1407_cast_fp16")]; + tensor var_1410_begin_0 = const()[name = tensor("op_1410_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1410_end_0 = const()[name = tensor("op_1410_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1410_end_mask_0 = const()[name = tensor("op_1410_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1410_squeeze_mask_0 = const()[name = tensor("op_1410_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1410_cast_fp16 = slice_by_index(begin = var_1410_begin_0, end = var_1410_end_0, end_mask = var_1410_end_mask_0, squeeze_mask = var_1410_squeeze_mask_0, x = var_1407_cast_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1425_begin_0 = const()[name = tensor("op_1425_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_1425_end_0 = const()[name = tensor("op_1425_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_1425_end_mask_0 = const()[name = tensor("op_1425_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1425_cast_fp16 = slice_by_index(begin = var_1425_begin_0, end = var_1425_end_0, end_mask = var_1425_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1428_begin_0 = const()[name = tensor("op_1428_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1428_end_0 = const()[name = tensor("op_1428_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1428_end_mask_0 = const()[name = tensor("op_1428_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1428_squeeze_mask_0 = const()[name = tensor("op_1428_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1428_cast_fp16 = slice_by_index(begin = var_1428_begin_0, end = var_1428_end_0, end_mask = var_1428_end_mask_0, squeeze_mask = var_1428_squeeze_mask_0, x = var_1425_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1443_begin_0 = const()[name = tensor("op_1443_begin_0"), val = tensor([0, 14, 0, 0])]; + tensor var_1443_end_0 = const()[name = tensor("op_1443_end_0"), val = tensor([1, 15, 1, 1500])]; + tensor var_1443_end_mask_0 = const()[name = tensor("op_1443_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_1443_cast_fp16 = slice_by_index(begin = var_1443_begin_0, end = var_1443_end_0, end_mask = var_1443_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1443_cast_fp16")]; + tensor var_1446_begin_0 = const()[name = tensor("op_1446_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1446_end_0 = const()[name = tensor("op_1446_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_1446_end_mask_0 = const()[name = tensor("op_1446_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_1446_squeeze_mask_0 = const()[name = tensor("op_1446_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_1446_cast_fp16 = slice_by_index(begin = var_1446_begin_0, end = var_1446_end_0, end_mask = var_1446_end_mask_0, squeeze_mask = var_1446_squeeze_mask_0, x = var_1443_cast_fp16)[name = tensor("op_1446_cast_fp16")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor(1)]; + tensor var_1454_interleave_0 = const()[name = tensor("op_1454_interleave_0"), val = tensor(false)]; + tensor var_1454_cast_fp16 = concat(axis = var_1453, interleave = var_1454_interleave_0, values = (var_1356_cast_fp16, var_1374_cast_fp16, var_1392_cast_fp16, var_1410_cast_fp16, var_1428_cast_fp16, var_1446_cast_fp16))[name = tensor("op_1454_cast_fp16")]; + tensor obj_axes_0 = const()[name = tensor("obj_axes_0"), val = tensor([1])]; + tensor obj_keep_dims_0 = const()[name = tensor("obj_keep_dims_0"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_1454_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file