# ------------------------------------------------------------------------- # Copyright (c) Microsoft Corporation. All rights reserved. # Licensed under the MIT License. # ------------------------------------------------------------------------- """ This converts GPT2 or T5 model to onnx with beam search operator. Example 1: convert gpt2 model with beam search: python convert_generation.py -m gpt2 --output gpt2_beam_search.onnx Example 2: convert gpt2 model with beam search containing specific cuda optimizations: python convert_generation.py -m gpt2 --output gpt2_beam_search.onnx --use_gpu \ --past_present_share_buffer --use_decoder_masked_attention Example 3: convert gpt2 model with beam search with mixed precision and enable SkipLayerNorm strict mode: python convert_generation.py -m gpt2 --output gpt2_beam_search.onnx --use_gpu -p fp16 --use_sln_strict_mode Example 4: convert T5 model with beam search in two steps: cd ./models/t5 python convert_to_onnx.py -m t5-small cd ../.. python convert_generation.py -m t5-small --model_type t5 \ --decoder_onnx ./models/t5/onnx_models/t5-small_decoder.onnx \ --encoder_decoder_init_onnx ./models/t5/onnx_models/t5-small_encoder_decoder_init.onnx \ --output ./models/t5/onnx_models/t5_small_beam_search.onnx Example 5: convert T5 model with beam search. All in one step: python convert_generation.py -m t5-small --model_type t5 --output ./models/t5/onnx_models/t5_small_beam_search.onnx Example 6: convert T5 model with beam search containing specific cuda optimizations. All in one step: python convert_generation.py -m t5-small --model_type t5 --output ./models/t5/onnx_models/t5_small_beam_search.onnx \ --use_gpu --past_present_share_buffer --use_decoder_masked_attention Example 7: convert MT5 model with external data file like mt5-base-beamsearch.onnx.data in below example. python convert_generation.py -m google/mt5-base --model_type mt5 --output mt5-base-beamsearch.onnx -e Example 8: convert gpt2 model with greedy search: python convert_generation.py -m gpt2 --output gpt2_greedy_search.onnx --num_beams 1 --num_return_sequences 1 Example 9: convert gpt2 model with sampling: python convert_generation.py -m gpt2 --output gpt2_sampling.onnx --num_beams 1 --num_return_sequences 1 --top_p 0.6 """ import argparse import logging import math import os import time from enum import Enum from pathlib import Path from typing import Any, Dict, List, Optional, Union import numpy as np import onnx import torch from benchmark_helper import Precision, setup_logger from fusion_utils import NumpyHelper from onnx import GraphProto, ModelProto, TensorProto from onnx_model import OnnxModel from transformers import ( GPT2Config, GPT2LMHeadModel, GPT2Tokenizer, MT5Config, MT5ForConditionalGeneration, T5Config, T5ForConditionalGeneration, T5Tokenizer, ) from onnxruntime import GraphOptimizationLevel, InferenceSession, SessionOptions, get_available_providers from onnxruntime.transformers.models.gpt2.convert_to_onnx import main as convert_gpt2_to_onnx from onnxruntime.transformers.models.gpt2.gpt2_helper import PRETRAINED_GPT2_MODELS from onnxruntime.transformers.models.t5.convert_to_onnx import export_onnx_models as export_t5_onnx_models from onnxruntime.transformers.models.t5.t5_helper import PRETRAINED_MT5_MODELS, PRETRAINED_T5_MODELS logger = logging.getLogger("") class GenerationType(Enum): BEAMSEARCH = "beam_search" GREEDYSEARCH = "greedy_search" SAMPLING = "sampling" def __str__(self): return self.value def parse_arguments(argv: Optional[List[str]] = None) -> argparse.Namespace: """Parse arguments Args: argv (Optional[List[str]], optional): _description_. Defaults to None. Returns: argparse.Namespace: Parsed arguments. """ parser = argparse.ArgumentParser() input_group = parser.add_argument_group("Input options") input_group.add_argument( "-m", "--model_name_or_path", required=True, type=str, help="Pytorch model checkpoint path, or pretrained model name in the list: " + ", ".join(PRETRAINED_GPT2_MODELS + PRETRAINED_T5_MODELS + PRETRAINED_MT5_MODELS), ) input_group.add_argument( "--model_type", required=False, type=str, default="gpt2", choices=["gpt2", "t5", "mt5"], help="Model type (default is gpt2) in the list: " + ", ".join(["gpt2", "t5", "mt5"]), ) input_group.add_argument( "--cache_dir", required=False, type=str, default=os.path.join(".", "cache_models"), help="Directory to cache pre-trained models", ) input_group.add_argument( "--decoder_onnx", required=False, type=str, default="", help="Path of onnx model for decoder. Specify it when you have exported the model.", ) input_group.add_argument( "--encoder_decoder_init_onnx", required=False, type=str, default="", help="Path of ONNX model for encoder and decoder initialization. Specify it when you have exported the model.", ) parser.add_argument( "--verbose", required=False, action="store_true", help="Print more information", ) parser.set_defaults(verbose=False) output_group = parser.add_argument_group("Output options") output_group.add_argument( "--output", required=True, type=str, help="Output path for onnx model with beam search.", ) output_group.add_argument( "-p", "--precision", required=False, type=Precision, default=Precision.FLOAT32, choices=[Precision.FLOAT32, Precision.FLOAT16], help="Precision of model to run. fp32 for full precision, fp16 for half or mixed precision", ) output_group.add_argument( "-b", "--op_block_list", required=False, nargs="*", default=["auto"], help="Disable certain onnx operators when exporting model to onnx format. When using default" 'value for gpt2 type of model fp16 precision, it will be set to ["Add", "LayerNormalization",' ' "SkipLayerNormalization", "FastGelu"]. Other situation, it will be set to []', ) output_group.add_argument( "-e", "--use_external_data_format", required=False, action="store_true", help="save external data for model > 2G", ) output_group.set_defaults(use_external_data_format=False) output_group.add_argument( "-s", "--run_shape_inference", required=False, action="store_true", help="run shape inference" ) output_group.set_defaults(run_shape_inference=False) output_group.add_argument( "-dpvs", "--disable_pad_vocab_size", required=False, action="store_true", help="Do not pad logits MatMul weight to be a multiple of 8 along the dimension where dim value is" " the vocab size. The logits MatMul may hence be of poor performance for fp16 precision.", ) output_group.set_defaults(disable_pad_vocab_size=False) output_group.add_argument( "-dsgd", "--disable_separate_gpt2_decoder_for_init_run", required=False, action="store_true", help="Do not create separate decoder subgraphs for initial and remaining runs. This does not allow " "for optimizations based on sequence lengths in each subgraph", ) output_group.set_defaults(disable_separate_gpt2_decoder_for_init_run=False) output_group.add_argument( "-i", "--disable_shared_initializers", required=False, action="store_true", help="do not share initializers in encoder and decoder for T5 or in the init decoder and decoder for " "GPT2. It will increase memory usage of t5/mt5/gpt2 models.", ) output_group.set_defaults(disable_shared_initializers=False) model_group = parser.add_argument_group("Beam search parameters that stored in the output model") model_group.add_argument( "--output_sequences_scores", required=False, action="store_true", help="output sequences scores", ) model_group.set_defaults(output_sequences_scores=False) model_group.add_argument( "--output_token_scores", required=False, action="store_true", help="output token scores", ) model_group.set_defaults(output_token_scores=False) model_group.add_argument("--early_stopping", required=False, action="store_true") model_group.set_defaults(early_stopping=False) model_group.add_argument( "--no_repeat_ngram_size", type=int, required=False, default=0, help="No repeat ngram size", ) model_group.add_argument( "--vocab_mask", required=False, action="store_true", help="Enable vocab_mask. This mask applies only to every generated token to filter some bad words.", ) model_group.set_defaults(vocab_mask=False) model_group.add_argument( "--past_present_share_buffer", required=False, action="store_true", help="Use shared buffer for past and present, currently work for gpt2 greedy/sampling search.", ) model_group.set_defaults(past_present_share_buffer=False) model_group.add_argument( "--use_decoder_masked_attention", required=False, action="store_true", help="Uses `DecoderMaskedSelfAttention` or `DecoderMaskedMultiHeadAttention` to optimize the decoding Attention computation. " "Must be used with `past_present_share_buffer`. Currently, only Attention head sizes of 32, 64 and 128 are supported.", ) model_group.set_defaults(use_decoder_masked_attention=False) model_group.add_argument( "--prefix_vocab_mask", required=False, action="store_true", help="Enable prefix_vocab_mask. This mask can be used to filter bad words in the first generated token only", ) model_group.set_defaults(prefix_vocab_mask=False) model_group.add_argument( "--custom_attention_mask", required=False, action="store_true", help="Enable custom_attention_mask. This mask can be used to replace default encoder attention mask", ) model_group.set_defaults(custom_attention_mask=False) model_group.add_argument( "--presence_mask", required=False, action="store_true", help="Presence mask for custom sampling", ) model_group.set_defaults(presence_mask=False) model_group.add_argument( "--seed", required=False, action="store_true", help="Random seed for sampling op", ) model_group.set_defaults(seed=False) beam_parameters_group = parser.add_argument_group( "Beam search parameters not stored in the output model, for testing parity and performance" ) beam_parameters_group.add_argument("--min_length", type=int, required=False, default=1, help="Min sequence length") beam_parameters_group.add_argument("--max_length", type=int, required=False, default=50, help="Max sequence length") beam_parameters_group.add_argument("--num_beams", type=int, required=False, default=4, help="Beam size") beam_parameters_group.add_argument( "--num_return_sequences", type=int, required=False, default=1, help="Number of return sequence <= num_beams", ) beam_parameters_group.add_argument( "--length_penalty", type=float, required=False, default=1, help="Positive. >1 to penalize and <1 to encourage short sentence.", ) beam_parameters_group.add_argument( "--repetition_penalty", type=float, required=False, default=1, help="Positive. >1 to penalize and <1 to encourage.", ) beam_parameters_group.add_argument( "--temperature", type=float, required=False, default=1.0, help="The value used to module the next token probabilities.", ) beam_parameters_group.add_argument( "--top_p", type=float, required=False, default=1.0, help="Top P for sampling", ) beam_parameters_group.add_argument( "--filter_value", type=float, required=False, default=-float("Inf"), help="Filter value for Top P sampling", ) beam_parameters_group.add_argument( "--min_tokens_to_keep", type=int, required=False, default=1, help="Minimum number of tokens we keep per batch example in the output.", ) beam_parameters_group.add_argument( "--presence_penalty", type=float, required=False, default=0.0, help="presence penalty for custom sampling.", ) beam_parameters_group.add_argument( "--custom", type=int, required=False, default=0, help="If 1 customized top P logic is applied", ) beam_parameters_group.add_argument( "--vocab_size", type=int, required=False, default=-1, help="Vocab_size of the underlying model used to decide the shape of vocab mask", ) beam_parameters_group.add_argument( "--eos_token_id", type=int, required=False, default=-1, help="custom eos_token_id for generating model with existing onnx encoder/decoder", ) beam_parameters_group.add_argument( "--pad_token_id", type=int, required=False, default=-1, help="custom pad_token_id for generating model with existing onnx encoder/decoder", ) test_group = parser.add_argument_group("Other options for testing parity and performance") test_group.add_argument( "--use_sln_strict_mode", required=False, action="store_true", help="Enable strict mode for SLN in CUDA provider. This ensures a better accuracy but will be slower.", ) test_group.set_defaults(use_sln_strict_mode=False) test_group.add_argument( "--use_gpu", required=False, action="store_true", help="use GPU for inference. Required for fp16." ) test_group.set_defaults(use_gpu=False) test_group.add_argument( "--disable_parity", required=False, action="store_true", help="do not run parity test", ) test_group.set_defaults(disable_parity=False) test_group.add_argument( "--disable_perf_test", required=False, action="store_true", help="do not run perf test", ) test_group.set_defaults(disable_perf_test=False) test_group.add_argument( "--torch_performance", required=False, action="store_true", help="test PyTorch performance", ) test_group.set_defaults(torch_performance=False) test_group.add_argument( "--total_runs", required=False, type=int, default=1, help="Number of times of inference for latency measurement", ) test_group.add_argument( "--save_test_data", required=False, action="store_true", help="save test data for onnxruntime_perf_test tool", ) test_group.set_defaults(save_test_data=False) args = parser.parse_args(argv) return args def gpt2_to_onnx(args: argparse.Namespace): """Convert GPT-2 model to onnx Args: args (argparse.Namespace): arguments parsed from command line """ model_name = args.model_name_or_path arguments = [ "--model_name_or_path", model_name, "--output", args.decoder_onnx, "--optimize_onnx", "--precision", "fp32" if args.precision == Precision.FLOAT32 else "fp16", "--test_runs", "1", "--test_cases", "10", "--overwrite", # Overwrite onnx file if existed ] if args.cache_dir: arguments.extend(["--cache_dir", args.cache_dir]) if args.use_gpu: arguments.append("--use_gpu") if args.use_external_data_format: arguments.append("--use_external_data_format") if len(args.op_block_list): arguments.extend(["--op_block_list"]) arguments.extend(args.op_block_list) if args.precision == Precision.FLOAT16: assert args.use_gpu, "fp16 or mixed precision model cannot run in CPU. Please add --use_gpu" # TODO(tianleiwu): Use auto mixed precision for fp16 conversion: arguments.append('--auto_mixed_precision') # Need change cuda kernel to support a combination of fp32 logits and fp16 past state. # Currently logits and past state shall be same data type. if args.verbose: logger.info(f"arguments for convert_to_onnx:{arguments}") convert_gpt2_to_onnx(argv=arguments) def t5_to_onnx(args: argparse.Namespace): """Convert T5 model to onnx Args: args (argparse.Namespace): arguments parsed from command line """ paths = export_t5_onnx_models( args.model_name_or_path, args.cache_dir, Path(args.output).parent, use_gpu=args.use_gpu, use_external_data_format=args.use_external_data_format, optimize_onnx=(args.precision != Precision.FLOAT16), precision=args.precision, verbose=False, use_decoder_start_token=False, merge_encoder_and_decoder_init=True, overwrite=True, disable_auto_mixed_precision=False, use_int32_inputs=True, model_type=args.model_type, ) logger.debug(f"onnx model for encoder: {paths[0]}") logger.debug(f"onnx model for decoder: {paths[1]}") args.encoder_decoder_init_onnx = paths[0] args.decoder_onnx = paths[1] def shape_inference(onnx_path: str, use_external_data_format: bool = True): """Shape inference on an onnx file, which will be overwritten. Args: onnx_path (str): Path of onnx model use_external_data_format(bool): output tensors to external data or not. """ # Run symbolic shape inference to walk around ORT shape inference issue for subgraph. from onnxruntime.tools.symbolic_shape_infer import SymbolicShapeInference model = onnx.load_model(onnx_path, load_external_data=True) out = SymbolicShapeInference.infer_shapes(model, auto_merge=True, guess_output_rank=False) if out: OnnxModel.save(out, onnx_path, save_as_external_data=use_external_data_format) else: logger.warning("Failed to run symbolic shape inference on the model.") def pad_weights_of_logits_matmul(onnx_path: str, use_external_data_format: bool = True) -> bool: """Pad the logits MatMul weight in the provided decoder model, which will be overwritten. Args: onnx_path (str): Path of onnx model use_external_data_format(bool): output tensors to external data or not. """ decoder_model_proto = onnx.load_model(onnx_path, load_external_data=True) logits_output_name = decoder_model_proto.graph.output[0].name decoder_model = OnnxModel(decoder_model_proto) output_name_to_node = decoder_model.output_name_to_node() assert logits_output_name in output_name_to_node matmul_node = output_name_to_node[logits_output_name] # Sanity check - the logits need to be produced by a MatMul node if matmul_node.op_type != "MatMul": return False # The logits MatMul weight MUST be an initializer (or) # it MUST be flowing through a Transpose whose input is # an initializer pad_along_axis_1 = True logits_weight = decoder_model.get_initializer(matmul_node.input[1]) if logits_weight is None: transpose_before_matmul = decoder_model.match_parent(matmul_node, "Transpose", 1) if transpose_before_matmul is None: return False logits_weight = decoder_model.get_initializer(transpose_before_matmul.input[0]) if logits_weight is None: return False pad_along_axis_1 = False # The logits MatMul weight MUST be fp16 if logits_weight.data_type != TensorProto.DataType.FLOAT16: return False # The logits MatMul weight MUST be 2-dimensional if len(logits_weight.dims) != 2: return False # Pad and over-write the initializer (if needed) actual_vocab_size = logits_weight.dims[1] if (actual_vocab_size % 8) == 0: # Already "padded" return True padded_vocab_size = math.ceil(actual_vocab_size / 8) * 8 padding = padded_vocab_size - actual_vocab_size # TODO(hasesh): Handle cases where the fp16 data is stored in the # non-raw data field if logits_weight.raw_data: if pad_along_axis_1: padding_data = np.zeros((logits_weight.dims[0], padding), dtype=np.float16) weight_with_padding = np.concatenate((NumpyHelper.to_array(logits_weight), padding_data), axis=1) logits_weight.dims[1] = padded_vocab_size else: padding_data = np.zeros((padding, logits_weight.dims[1]), dtype=np.float16) weight_with_padding = np.concatenate((NumpyHelper.to_array(logits_weight), padding_data), axis=0) logits_weight.dims[0] = padded_vocab_size logits_weight.raw_data = weight_with_padding.tobytes() else: return False # Save the model OnnxModel.save(decoder_model_proto, onnx_path, save_as_external_data=use_external_data_format) return True def create_ort_session(model_path: str, use_gpu: bool, use_sln_strict_mode: bool) -> InferenceSession: """Create OnnxRuntime session. Args: model_path (str): onnx model path use_gpu (bool): use GPU or not use_sln_strict_mode (bool): use strict mode for skip layer normalization or not Raises: RuntimeError: CUDAExecutionProvider is not available when --use_gpu is specified. Returns: onnxruntime.InferenceSession: The created session. """ sess_options = SessionOptions() sess_options.graph_optimization_level = GraphOptimizationLevel.ORT_DISABLE_ALL execution_providers = ["CUDAExecutionProvider", "CPUExecutionProvider"] if use_gpu else ["CPUExecutionProvider"] if use_gpu: if "CUDAExecutionProvider" not in get_available_providers(): raise RuntimeError("CUDAExecutionProvider is not available for --use_gpu!") else: logger.info("use CUDAExecutionProvider") if use_sln_strict_mode: cuda_provider_options = {"enable_skip_layer_norm_strict_mode": True} provider_options = {"CUDAExecutionProvider": cuda_provider_options} execution_providers = [ (name, provider_options[name]) if name in provider_options else name for name in execution_providers ] ort_session = InferenceSession(model_path, sess_options, providers=execution_providers) return ort_session def verify_gpt2_subgraph(graph: onnx.GraphProto, precision: Precision): """Verify GPT-2 subgraph Args: graph (onnx.GraphProto): onnx graph of GPT-2 precision (Precision): Precision (FLOAT16 or FLOAT32) of the model. Raises: ValueError: Number of inputs not expected. ValueError: Input name is not expected. ValueError: Input data type is not expected. ValueError: Number of outputs not expected. ValueError: Output name is not expected. ValueError: Output data type is not expected. """ is_float16 = precision == Precision.FLOAT16 input_count = len(graph.input) layer_count = input_count - 3 assert layer_count >= 1 expected_inputs = ["input_ids", "position_ids", "attention_mask"] + [f"past_{i}" for i in range(layer_count)] if len(graph.input) != len(expected_inputs): raise ValueError(f"Number of inputs expected to be {len(expected_inputs)}. Got {len(graph.input)}") for i, expected_input in enumerate(expected_inputs): if graph.input[i].name != expected_input: raise ValueError(f"Input {i} is expected to be {expected_input}. Got {graph.input[i].name}") expected_type = TensorProto.INT32 if i >= 3: expected_type = TensorProto.FLOAT16 if is_float16 else TensorProto.FLOAT input_type = graph.input[i].type.tensor_type.elem_type if input_type != expected_type: raise ValueError(f"Input {i} is expected to have onnx data type {expected_type}. Got {input_type}") logger.info("Verifying GPT-2 graph inputs: name and data type are good.") expected_outputs = ["logits"] + [f"present_{i}" for i in range(layer_count)] if len(graph.output) != len(expected_outputs): raise ValueError(f"Number of outputs expected to be {len(expected_outputs)}. Got {len(graph.output)}") for i, expected_output in enumerate(expected_outputs): if graph.output[i].name != expected_output: raise ValueError(f"Output {i} is expected to be {expected_output}. Got {graph.output[i].name}") expected_type = TensorProto.FLOAT16 if is_float16 else TensorProto.FLOAT output_type = graph.output[i].type.tensor_type.elem_type if output_type != expected_type: raise ValueError(f"Input {i} is expected to have onnx data type {expected_type}. Got {output_type}") logger.info("Verifying GPT-2 graph outputs: name and data type are good.") # TODO(tianleiwu): verify shapes of inputs and outputs. return def verify_t5_decoder_subgraph(graph: onnx.GraphProto, precision: Precision): """Verify T5 decoder subgraph Args: graph (onnx.GraphProto): onnx graph of T5 decoder precision (Precision): Precision (FLOAT16 or FLOAT32) of the model. Raises: ValueError: Number of inputs not expected. ValueError: Input name is not expected. ValueError: Input data type is not expected. ValueError: Number of outputs not expected. ValueError: Output name is not expected. ValueError: Output data type is not expected. """ is_float16 = precision == Precision.FLOAT16 float_type = TensorProto.FLOAT16 if is_float16 else TensorProto.FLOAT input_count = len(graph.input) layer_count = (input_count - 2) // 4 assert layer_count >= 1 # Expect inputs: # input_ids: int32 (B, 1) # encoder_attention_mask: int32 (B, encode_sequence_length) # past_key_self_0: (B, num_heads, past_decode_sequence_length, head_size) # past_value_self_0: (B, num_heads, past_decode_sequence_length, head_size) # ... (for each self attention layer) # past_key_cross_0: (B, num_heads, encode_sequence_length, head_size) # past_value_cross_0: (B, num_heads, encode_sequence_length, head_size) # ... (for each cross attention layer) # TODO: encoder_hidden_states is optional expected_inputs = ["input_ids", "encoder_attention_mask"] for i in range(layer_count): expected_inputs.append(f"past_key_self_{i}") expected_inputs.append(f"past_value_self_{i}") for i in range(layer_count): expected_inputs.append(f"past_key_cross_{i}") expected_inputs.append(f"past_value_cross_{i}") if len(graph.input) != len(expected_inputs): raise ValueError(f"Number of inputs expected to be {len(expected_inputs)}. Got {len(graph.input)}") for i, expected_input in enumerate(expected_inputs): if graph.input[i].name != expected_input: raise ValueError(f"Input {i} is expected to be {expected_input}. Got {graph.input[i].name}") expected_type = TensorProto.INT32 if i < 2 else float_type input_type = graph.input[i].type.tensor_type.elem_type if input_type != expected_type: raise ValueError(f"Input {i} is expected to have onnx data type {expected_type}. Got {input_type}") # Expect outputs: # logits: (B, 1, vocab_size) # present_key_self_0: (B, num_heads, past_decode_sequence_length + 1, head_size) # present_value_self_0: (B, num_heads, past_decode_sequence_length + 1, head_size) # ... (for each self attention layer) expected_outputs = ["logits"] for i in range(layer_count): expected_outputs.append(f"present_key_self_{i}") expected_outputs.append(f"present_value_self_{i}") if len(graph.output) != len(expected_outputs): raise ValueError(f"Number of outputs expected to be {len(expected_outputs)}. Got {len(graph.output)}") for i, expected_output in enumerate(expected_outputs): if graph.output[i].name != expected_output: raise ValueError(f"Output {i} is expected to be {expected_output}. Got {graph.output[i].name}") output_type = graph.output[i].type.tensor_type.elem_type if output_type != float_type: raise ValueError(f"Output {i} is expected to have onnx data type {float_type}. Got {output_type}") def verify_t5_encoder_decoder_init_subgraph(graph: onnx.GraphProto, precision: Precision): """Verify T5 decoder subgraph Args: graph (onnx.GraphProto): onnx graph of T5 decoder precision (Precision): Precision (FLOAT16 or FLOAT32) of the model. Raises: ValueError: Number of inputs not expected. ValueError: Input name is not expected. ValueError: Input data type is not expected. ValueError: Number of outputs not expected. ValueError: Output name is not expected. ValueError: Output data type is not expected. """ is_float16 = precision == Precision.FLOAT16 layer_count = (len(graph.output) - 2) // 4 assert layer_count >= 1 # Expect 3 inputs: # encoder_input_ids: int32 (B, encode_sequence_length) # encoder_attention_mask: int32 (B, encode_sequence_length) # decoder_input_ids: int32 (B, 1) expected_inputs = ["encoder_input_ids", "encoder_attention_mask", "decoder_input_ids"] if len(graph.input) != len(expected_inputs): raise ValueError(f"Number of inputs expected to be {len(expected_inputs)}. Got {len(graph.input)}") for i, expected_input in enumerate(expected_inputs): if graph.input[i].name != expected_input: raise ValueError(f"Input {i} is expected to be {expected_input}. Got {graph.input[i].name}") expected_type = TensorProto.INT32 input_type = graph.input[i].type.tensor_type.elem_type if input_type != expected_type: raise ValueError(f"Input {i} is expected to have onnx data type {expected_type}. Got {input_type}") # Expected outputs: # logits: (B, 1, vocab_size) # encoder_hidden_states: (B, encode_sequence_length, encoder_hidden_size) # present_key_self_0: (B, num_heads, 1, head_size) # present_value_self_0: (B, num_heads, 1, head_size) # ... (for each self attention layer) # present_key_cross_0: (B, num_heads, encode_sequence_length, head_size) # present_value_cross_0: (B, num_heads, encode_sequence_length, head_size) # ... (for each cross attention layer) expected_outputs = ["logits", "encoder_hidden_states"] for i in range(layer_count): expected_outputs.append(f"present_key_self_{i}") expected_outputs.append(f"present_value_self_{i}") for i in range(layer_count): expected_outputs.append(f"present_key_cross_{i}") expected_outputs.append(f"present_value_cross_{i}") if len(graph.output) != len(expected_outputs): raise ValueError(f"Number of outputs expected to be {len(expected_outputs)}. Got {len(graph.output)}") for i, expected_output in enumerate(expected_outputs): if graph.output[i].name != expected_output: raise ValueError(f"Output {i} is expected to be {expected_output}. Got {graph.output[i].name}") expected_type = TensorProto.FLOAT16 if is_float16 else TensorProto.FLOAT output_type = graph.output[i].type.tensor_type.elem_type if output_type != expected_type: raise ValueError(f"Output {i} is expected to have onnx data type {expected_type}. Got {output_type}") logger.info("T5 encoder graph verified: name and data type of inputs and outputs are good.") def remove_shared_initializers( graph1: GraphProto, graph2: GraphProto, shared_prefix: str = "shared_", min_elements: int = 1024, signature_cache1: Optional[dict] = None, signature_cache2: Optional[dict] = None, ): """Remove initializers with same value from two graphs. Args: graph1 (GraphProto): the first graph to process graph2 (GraphProto): the second graph to process shared_prefix (str): add prefix to the shared initializers among two graphs min_elements (int, optional): minimal number of elements for initializers to be considered. Defaults to 1024. signature_cache1 (dict): Optional dictionary to store data signatures of tensors in graph1 in order to speed up comparison signature_cache2 (dict): Optional dictionary to store data signatures of tensors in graph2 in order to speed up comparison """ mapping_initializers_1 = {} mapping_initializers_2 = {} shared_initializers_1 = [] shared_initializers_2 = [] shared_initializers_names = [] for initializer1 in graph1.initializer: if not (initializer1.dims and sum(initializer1.dims) >= min_elements): continue for initializer2 in graph2.initializer: if not (initializer2.dims and sum(initializer2.dims) >= min_elements): continue if OnnxModel.has_same_value(initializer1, initializer2, signature_cache1, signature_cache2): mapping_initializers_1[initializer1.name] = shared_prefix + initializer2.name shared_initializers_1.append(initializer1) if initializer2.name not in mapping_initializers_2: shared_name = shared_prefix + initializer2.name mapping_initializers_2[initializer2.name] = shared_name shared_initializers_2.append(initializer2) shared_initializers_names.append(shared_name) break logger.debug(f"shared initializers:{shared_initializers_names}") # Make sure new name does not exist in graph 1 for node in graph1.node: for j in range(len(node.input)): if node.input[j] in shared_initializers_names: raise RuntimeError(f"name is found in graph 1: {node.input[j]}") # Make sure new name does not exist in graph 2 for node in graph2.node: for j in range(len(node.input)): if node.input[j] in shared_initializers_names: raise RuntimeError(f"name is found in graph 2: {node.input[j]}") # Remove shared initializers from graph 2 for initializer in shared_initializers_2: graph2.initializer.remove(initializer) # Rename value info for old names in graph 2 for value_info in graph2.value_info: if value_info.name in mapping_initializers_2: value_info.name = mapping_initializers_2[value_info.name] # Rename nodes inputs in graph 2: for node in graph2.node: for j in range(len(node.input)): if node.input[j] in mapping_initializers_2: new_name = mapping_initializers_2[node.input[j]] logger.debug(f"graph 2 rename node {node.name} input {j} from {node.input[j]} to {new_name}") node.input[j] = new_name # Remove shared initializers from graph 1 for initializer in shared_initializers_1: graph1.initializer.remove(initializer) # Rename value info for old names in graph 1 for value_info in graph1.value_info: if value_info.name in mapping_initializers_1: value_info.name = mapping_initializers_1[value_info.name] # Rename nodes inputs in graph 1: for node in graph1.node: for j in range(len(node.input)): if node.input[j] in mapping_initializers_1: new_name = mapping_initializers_1[node.input[j]] logger.debug(f"graph 1 rename node {node.name} input {j} from {node.input[j]} to {new_name}") node.input[j] = new_name # Rename shared initializers in graph 2 for initializer in shared_initializers_2: initializer.name = mapping_initializers_2[initializer.name] for initializer in shared_initializers_2: shape = onnx.numpy_helper.to_array(initializer).shape value_info = onnx.helper.make_tensor_value_info(initializer.name, initializer.data_type, shape) # Need add value_info for initializers moved to parent graph. Otherwise, ORT will fail. graph1.value_info.append(value_info) graph2.value_info.append(value_info) return shared_initializers_2 def get_shared_initializers(encoder_model: ModelProto, decoder_model: ModelProto): encoder = OnnxModel(encoder_model) decoder = OnnxModel(decoder_model) encoder.add_prefix_to_names("e_") decoder.add_prefix_to_names("d_") signature_cache1, signature_cache2 = {}, {} encoder.remove_duplicated_initializer(signature_cache1) decoder.remove_duplicated_initializer(signature_cache2) initializers = remove_shared_initializers( decoder.model.graph, encoder.model.graph, shared_prefix="s_", signature_cache1=signature_cache1, signature_cache2=signature_cache2, ) return initializers def move_initializers( graph: GraphProto, min_elements: int = 1024, ) -> List[TensorProto]: """Remove initializers of a graph, when they have number of elements larger than a threshold. Args: graph (GraphProto): the graph. min_elements (int, optional): minimal number of elements for initializers to be considered. Defaults to 1024. Returns: List[TensorProto]: initializers that are removed from the graph. """ moved_initializers = [] for tensor in graph.initializer: if not (tensor.dims and sum(tensor.dims) >= min_elements): continue moved_initializers.append(tensor) for initializer in moved_initializers: graph.initializer.remove(initializer) # Add type info, otherwise ORT will raise error: "input arg (*) does not have type information set by parent node." for initializer in moved_initializers: shape = onnx.numpy_helper.to_array(initializer).shape value_info = onnx.helper.make_tensor_value_info(initializer.name, initializer.data_type, shape) graph.value_info.append(value_info) return moved_initializers def _attribute_to_pair(attribute): """ Convert attribute to kwarg format for use with onnx.helper.make_node. :parameter attribute: attribute in AttributeProto format. :return: attribute in {key: value} format. """ if attribute.type == 0: raise ValueError(f"attribute {attribute.name} does not have type specified.") # Based on attribute type definitions from AttributeProto # definition in https://github.com/onnx/onnx/blob/master/onnx/onnx.proto if attribute.type == 1: value = attribute.f elif attribute.type == 2: value = attribute.i elif attribute.type == 3: value = attribute.s elif attribute.type == 4: value = attribute.t elif attribute.type == 5: value = attribute.g elif attribute.type == 6: value = attribute.floats elif attribute.type == 7: value = attribute.ints elif attribute.type == 8: value = attribute.strings elif attribute.type == 9: value = attribute.tensors elif attribute.type == 10: value = attribute.graphs else: raise ValueError(f"attribute {attribute.name} has unsupported type {attribute.type}.") return (attribute.name, value) def kwargs_of(node): kwargs = {} for attr in node.attribute: (key, value) = _attribute_to_pair(attr) kwargs.update({key: value}) if node.domain: kwargs.update({"domain": node.domain}) return kwargs def shape_of(vi): return tuple([d.dim_param if (d.dim_param) else d.dim_value for d in vi.type.tensor_type.shape.dim]) def update_decoder_subgraph_past_present_share_buffer(subg: GraphProto): input_past_0 = 3 output_past_0 = 1 new_inputs = [] for i, vi in enumerate(subg.input): if i >= input_past_0: shape = shape_of(vi) vi = onnx.helper.make_tensor_value_info( # noqa: PLW2901 vi.name, elem_type=vi.type.tensor_type.elem_type, shape=[shape[0], shape[1], shape[2], "max_seq_len", shape[4]], ) new_inputs.extend([vi]) new_inputs.extend([onnx.helper.make_tensor_value_info("past_sequence_length", onnx.TensorProto.INT32, shape=[1])]) subg.ClearField("input") subg.input.extend(new_inputs) new_outputs = [] for i, vi in enumerate(subg.output): if i >= output_past_0: shape = shape_of(vi) vi = onnx.helper.make_tensor_value_info( # noqa: PLW2901 vi.name, elem_type=vi.type.tensor_type.elem_type, shape=[shape[0], shape[1], shape[2], "max_seq_len", shape[4]], ) new_outputs.extend([vi]) subg.ClearField("output") subg.output.extend(new_outputs) new_nodes = [] for node in subg.node: if node.op_type == "Attention": kwargs = kwargs_of(node) kwargs.update({"past_present_share_buffer": 1}) nis = [] nis.extend(node.input) while len(nis) < 6: nis.extend([""]) if len(nis) < 7: nis.extend(["past_sequence_length"]) node = onnx.helper.make_node("Attention", nis, node.output, name=node.name, **kwargs) # noqa: PLW2901 new_nodes.extend([node]) subg.ClearField("node") subg.node.extend(new_nodes) return subg def update_decoder_subgraph_use_decoder_masked_attention( subg: GraphProto, is_beam_search: bool, switch_attention: bool ) -> bool: """Update the Attention nodes to DecoderMaskedSelfAttention. Args: subg (GraphProto): GraphProto of the decoder subgraph is_beam_search (bool): Boolean specifying if the sampling algo is BeamSearch switch_attention (bool): Boolean specifying if `Attention` is to be switched with `DecoderMaskedSelfAttention` """ if is_beam_search: new_inputs = [] for _i, vi in enumerate(subg.input): new_inputs.extend([vi]) # Add 2 BeamSearch specific inputs new_inputs.extend([onnx.helper.make_tensor_value_info("beam_width", onnx.TensorProto.INT32, shape=[1])]) new_inputs.extend( [ onnx.helper.make_tensor_value_info( "cache_indirection", onnx.TensorProto.INT32, shape=["batch_size", "beam_width", "max_seq_len"] ) ] ) subg.ClearField("input") subg.input.extend(new_inputs) if switch_attention: decoder_masked_attention_supported_attr = [ "past_present_share_buffer", "num_heads", "scale", "mask_filter_value", "domain", ] new_nodes = [] for node in subg.node: if node.op_type == "Attention": kwargs = kwargs_of(node) for k in kwargs.copy(): # The Attention operator does not support different qkv hidden sizes when past/present # input/output exists (GPT2 model). Hence, we should never run into this. # But, if we do, do not go ahead with the optimization. if k == "qkv_hidden_sizes": return False if k not in decoder_masked_attention_supported_attr: # Log the fact that we are removing certain attributes from the node # We don't need to log it for "unidirectional" as we are aware that # decoding attention kernels are unidirectional by definition. if k != "unidirectional": logger.warning( f"Removing attribute: {k} from Attention node while switching to DecoderMaskedSelfAttention" ) del kwargs[k] nis = [] nis.extend(node.input) # Add 2 BeamSearch specific inputs if is_beam_search: while len(nis) < 7: nis.extend([""]) if len(nis) < 8: nis.extend(["beam_width"]) if len(nis) < 9: nis.extend(["cache_indirection"]) node = onnx.helper.make_node( # noqa: PLW2901 "DecoderMaskedSelfAttention", nis, node.output, name=node.name, **kwargs ) new_nodes.extend([node]) subg.ClearField("node") subg.node.extend(new_nodes) return True def find_past_seq_len_usage(subg: GraphProto): """Correct graph which originally use dim of past_seq_len from input_ids's shape which is fixed to max_seq_len after shared past/present buffer Args: subg (GraphProto): GraphProto of the decoder subgraph return: tensor_names_to_rename : set of tensor names which is equal to past_sequence_length nodes_to_remove : list of node to remove """ tensor_names_to_rename = set() nodes_to_remove = [] graph_input_names = {inp.name: index for index, inp in enumerate(subg.input)} input_name_to_nodes = {} output_name_to_node = {} for node in subg.node: for input_name in node.input: if input_name: if input_name not in input_name_to_nodes: input_name_to_nodes[input_name] = [node] else: input_name_to_nodes[input_name].append(node) for output_name in node.output: if output_name: output_name_to_node[output_name] = node for node in subg.node: # find "Shape(past_key_self..) --> Gather(*, 2)" if node.op_type == "Gather": if not node.input[1] or not node.input[0]: continue shape_tensor_name, shape_index_name = (node.input[0], node.input[1]) ini_gather_indices = None for tensor in subg.initializer: if tensor.name == shape_index_name: ini_gather_indices = tensor break if ini_gather_indices is None: continue gather_indices_arr = onnx.numpy_helper.to_array(ini_gather_indices) if gather_indices_arr.size == 1 and gather_indices_arr.item() == 2 and node.input[0] in output_name_to_node: shape_node = output_name_to_node[shape_tensor_name] if ( shape_node.op_type == "Shape" and shape_node.input[0] and shape_node.input[0] in graph_input_names and ( shape_node.input[0].startswith("past_key_self_") or shape_node.input[0].startswith("past_value_self_") ) ): tensor_names_to_rename.add(node.output[0]) nodes_to_remove.append(node) if len(input_name_to_nodes[shape_node.output[0]]) == 1: nodes_to_remove.append(shape_node) return tensor_names_to_rename, nodes_to_remove def replace_mha_with_gqa( model: OnnxModel, attn_mask: str, kv_num_heads: int = 0, world_size: int = 1, window_size: int = -1 ): # Insert attention_mask subgraph to calculate shared inputs for all GroupQueryAttention nodes # # attention_mask # / \ # ReduceSum Shape # | | # Sub Gather # | | # seqlens_k total_sequence_length # | | # Cast to int32 Cast to int32 model.add_initializer( onnx.helper.make_tensor( name="one", data_type=TensorProto.INT64, dims=[1], vals=[1], ) ) reduce_sum_node = onnx.helper.make_node( "ReduceSum", inputs=[attn_mask, "one"], outputs=[attn_mask + "_row_sums"], name=model.create_node_name("ReduceSum"), ) sub_node = onnx.helper.make_node( "Sub", inputs=[attn_mask + "_row_sums", "one"], outputs=["seqlens_k_int64"], name=model.create_node_name("Sub"), ) seqlen_k_cast_node = onnx.helper.make_node( "Cast", inputs=["seqlens_k_int64"], outputs=["seqlens_k"], name=model.create_node_name("Cast"), to=TensorProto.INT32, ) shape_node = onnx.helper.make_node( "Shape", inputs=[attn_mask], outputs=[attn_mask + "_shape"], name=model.create_node_name("Shape"), ) gather_node = onnx.helper.make_node( "Gather", inputs=[attn_mask + "_shape", "one"], outputs=["total_seq_len_int64"], name=model.create_node_name("Gather"), axis=0, ) total_seqlen_cast_node = onnx.helper.make_node( "Cast", inputs=["total_seq_len_int64"], outputs=["total_seq_len"], name=model.create_node_name("Cast"), to=TensorProto.INT32, ) model.model.graph.node.extend( [reduce_sum_node, sub_node, seqlen_k_cast_node, shape_node, gather_node, total_seqlen_cast_node] ) # Replace MultiHeadAttention with GroupQueryAttention # # When replacing, fuse the following subgraph: # # root_input # / | \ # MatMul MatMul MatMul # | | | # Add Add Add (optional Adds) # | | | # RotEmb RotEmb | # \ | / # MultiHeadAttention # # to this new subgraph: # # root_input # | # PackedMatMul (if possible) # | # PackedAdd (if possible) # | # GroupQueryAttention # mha_nodes = list(filter(lambda node: node.op_type == "MultiHeadAttention", model.model.graph.node)) for idx, node in enumerate(mha_nodes): # Detect Q path to MHA q_path_1 = model.match_parent_path(node, ["RotaryEmbedding", "Add", "MatMul"], [0, 0, 0]) q_path_2 = model.match_parent_path(node, ["RotaryEmbedding", "MatMul"], [0, 0]) q_rotary, q_add, q_matmul = None, None, None if q_path_1 is not None: q_rotary, q_add, q_matmul = q_path_1 elif q_path_2 is not None: q_rotary, q_matmul = q_path_2 # Detect K path to MHA k_path_1 = model.match_parent_path(node, ["RotaryEmbedding", "Add", "MatMul"], [1, 0, 0]) k_path_2 = model.match_parent_path(node, ["RotaryEmbedding", "MatMul"], [1, 0]) k_rotary, k_add, k_matmul = None, None, None if k_path_1 is not None: k_rotary, k_add, k_matmul = k_path_1 elif k_path_2 is not None: k_rotary, k_matmul = k_path_2 # Detect V path to MHA v_path_1 = model.match_parent_path(node, ["Add", "MatMul"], [2, 0]) v_path_2 = model.match_parent_path(node, ["MatMul"], [2]) v_add, v_matmul = None, None if v_path_1 is not None: v_add, v_matmul = v_path_1 elif v_path_2 is not None: v_matmul = v_path_2[0] # Get `interleaved` attribute from RotaryEmbedding interleaved = 0 if q_rotary is not None and k_rotary is not None: for att in q_rotary.attribute: if att.name == "interleaved": interleaved = att.i # Get `num_heads` attribute from MHA num_heads = 0 for att in node.attribute: if att.name == "num_heads": num_heads = att.i # Check if root_input to Q/K/V paths is the same root_input_is_same = q_matmul.input[0] == k_matmul.input[0] and k_matmul.input[0] == v_matmul.input[0] # Check if Q/K/V paths all have bias or all don't have bias all_paths_have_bias = q_add is not None and k_add is not None and v_add is not None all_paths_have_no_bias = q_add is None and k_add is None and v_add is None # Make PackedMatMul node if possible q_input_to_attention, k_input_to_attention, v_input_to_attention = "", "", "" if root_input_is_same and (all_paths_have_bias or all_paths_have_no_bias): qw = NumpyHelper.to_array(model.get_initializer(q_matmul.input[1])) kw = NumpyHelper.to_array(model.get_initializer(k_matmul.input[1])) vw = NumpyHelper.to_array(model.get_initializer(v_matmul.input[1])) dim = qw.shape[-1] qkv_weight = np.stack((qw, kw, vw), axis=1).reshape(dim, 3 * dim) qkv_weight = onnx.numpy_helper.from_array(qkv_weight, name=f"QKV_Weight_{idx}") model.add_initializer(qkv_weight) packed_matmul_node = onnx.helper.make_node( "MatMul", inputs=[q_matmul.input[0], qkv_weight.name], outputs=[f"{qkv_weight.name}_output"], name=model.create_node_name("MatMul"), ) model.model.graph.node.extend([packed_matmul_node]) model.model.graph.node.remove(q_matmul) model.model.graph.node.remove(k_matmul) model.model.graph.node.remove(v_matmul) q_input_to_attention = packed_matmul_node.output[0] # Make PackedAdd node if possible if all_paths_have_bias: qb = NumpyHelper.to_array(model.get_initializer(q_add.input[1])) kb = NumpyHelper.to_array(model.get_initializer(k_add.input[1])) vb = NumpyHelper.to_array(model.get_initializer(v_add.input[1])) dim = qb.shape[-1] qkv_bias = np.stack((qb, kb, vb), axis=0).reshape(3 * dim) qkv_bias = onnx.numpy_helper.from_array(qkv_bias, name=f"QKV_Bias_{idx}") model.add_initializer(qkv_bias) packed_add_node = onnx.helper.make_node( "Add", inputs=[packed_matmul_node.output[0], qkv_bias.name], outputs=[f"{qkv_bias.name}_output"], ) model.model.graph.node.extend([packed_add_node]) model.model.graph.node.remove(q_add) model.model.graph.node.remove(k_add) model.model.graph.node.remove(v_add) q_input_to_attention = packed_add_node.output[0] else: q_input_to_attention = q_matmul.output[0] k_input_to_attention = k_matmul.output[0] v_input_to_attention = v_matmul.output[0] # Make GQA node gqa_node = onnx.helper.make_node( "GroupQueryAttention", inputs=[ q_input_to_attention, # query k_input_to_attention, # key v_input_to_attention, # value node.input[6], # past_key node.input[7], # past_value seqlen_k_cast_node.output[0], # seqlens_k (for attention mask) total_seqlen_cast_node.output[0], # total_seq_len (for attention mask) q_rotary.input[2] if q_rotary is not None else "", # cos_cache (for rotary embeddings) q_rotary.input[3] if q_rotary is not None else "", # sin_cache (for rotary embeddings) ], outputs=node.output, name=node.name.replace("MultiHeadAttention", "GroupQueryAttention"), domain="com.microsoft", num_heads=num_heads // world_size, kv_num_heads=num_heads // world_size if kv_num_heads == 0 else kv_num_heads // world_size, local_window_size=window_size, do_rotary=int(q_rotary is not None and k_rotary is not None), rotary_interleaved=interleaved, ) model.model.graph.node.remove(node) model.model.graph.node.extend([gqa_node]) if q_rotary is not None: model.model.graph.node.remove(q_rotary) if k_rotary is not None: model.model.graph.node.remove(k_rotary) return model def update_decoder_subgraph_output_cross_attention(subg: GraphProto): input_self_past_0 = 1 # w/wo attention mask, w/wo hidden_state graph_input_names = [gi.name for gi in subg.input] while input_self_past_0 < 3 and not graph_input_names[input_self_past_0].startswith("past"): input_self_past_0 += 1 output_self_present_0 = 1 num_layers = (len(subg.output) - output_self_present_0) // 2 input_cross_past_0 = 2 * num_layers + input_self_past_0 past_key_cross_inputs = {subg.input[layer * 2 + input_cross_past_0].name: layer for layer in range(num_layers)} print(f" --past_key_cross_inputs={past_key_cross_inputs}") input_past_key_cross_0_shape = shape_of(subg.input[input_cross_past_0]) print(f"past_key_cross_0_shape is {input_past_key_cross_0_shape}") batch_size_dim = input_past_key_cross_0_shape[0] num_heads_dim = input_past_key_cross_0_shape[1] cross_seq_len_dim = input_past_key_cross_0_shape[2] num_layer_output_qk = 0 for node in subg.node: if (node.op_type == "DecoderMaskedMultiHeadAttention") and (node.input[1] in past_key_cross_inputs): print(f" -- add cross QK output from: node: {node.name} with output: {node.output}") num_layer_output_qk += 1 layer = past_key_cross_inputs[node.input[1]] cross_attention_out_name = f"output_cross_qk_{layer}" appended_names = [""] * (3 - len(node.output)) appended_names.append(cross_attention_out_name) node.output.extend(appended_names) node.attribute.extend([onnx.helper.make_attribute("output_qk", 1)]) cross_attention = onnx.helper.make_tensor_value_info( cross_attention_out_name, TensorProto.FLOAT, [batch_size_dim, num_heads_dim, 1, cross_seq_len_dim] ) subg.output.extend([cross_attention]) if num_layer_output_qk != num_layers: raise ValueError(f"Did not add cross QK for all layers{num_layers} vs {num_layer_output_qk}") def update_decoder_subgraph_share_buffer_and_use_decoder_masked_mha(subg: ModelProto): input_self_past_0 = 1 # w/wo attention mask, w/wo hidden_state graph_input_names = [gi.name for gi in subg.input] while input_self_past_0 < 3 and not graph_input_names[input_self_past_0].startswith("past"): input_self_past_0 += 1 output_self_past_0 = 1 num_layers = int((len(subg.input) - input_self_past_0) / 4) input_cross_past_0 = 2 * num_layers + input_self_past_0 new_nodes = [] old_nodes = [] for node in subg.node: if node.op_type == "MultiHeadAttention": old_nodes.extend([node]) # If not all the MultiHeadAttention nodes are fused, this optimization is not applicable if len(old_nodes) < num_layers: return False # Redirect the RelativePositionBias node's input from past_key_self_0.shape[2] to past_sequence_length. # There is only one RelativePositionBias node in T5 decoder subgraph. rel_pos_bias_node = None for node in subg.node: if node.op_type == "RelativePositionBias": rel_pos_bias_node = node break decoder_masked_attention_supported_attr = [ "past_present_share_buffer", "num_heads", "scale", "mask_filter_value", "domain", ] target_squeezed_past_seq_name = "past_sequence_length_squeezed_int64" tensor_names_to_rename, nodes_to_remove = find_past_seq_len_usage(subg) if len(tensor_names_to_rename) > 0: for name_to_rename in tensor_names_to_rename: print(f"Found tensor name {name_to_rename} to be renamed to {target_squeezed_past_seq_name}") for nr in nodes_to_remove: print(f"Found node to removed: type:{nr.op_type}, name:{nr.name}") squeeze_node = onnx.helper.make_node( "Squeeze", ["past_sequence_length"], ["past_sequence_length_squeezed"], name="node_past_sequence_length_squeeze", ) cast_node = onnx.helper.make_node( "Cast", ["past_sequence_length_squeezed"], [target_squeezed_past_seq_name], name="node_past_sequence_length_squeeze_cast", to=TensorProto.INT64, ) new_nodes.extend([squeeze_node, cast_node]) for node in subg.node: if len(node.output) > 0 and rel_pos_bias_node is not None and node.output[0] == rel_pos_bias_node.input[1]: cast_node = onnx.helper.make_node( "Cast", ["past_sequence_length"], ["past_sequence_length_int64"], name="past_sequence_length_cast", to=TensorProto.INT64, ) node.input[1] = cast_node.output[0] new_nodes.extend([cast_node]) if node.op_type == "MultiHeadAttention": kwargs = kwargs_of(node) for k in kwargs.copy(): if k not in decoder_masked_attention_supported_attr: del kwargs[k] # note: This logic only apply to T5 model where there is no bias in Attention node. nis = [ node.input[0], # query node.input[1], # key node.input[2], # value ] nis.extend([node.input[4] if len(node.input) > 4 else ""]) # 2D mask nis.extend([node.input[5] if len(node.input) > 5 else ""]) # attention_bias nis.extend([node.input[6] if len(node.input) > 6 else ""]) # past_key nis.extend([node.input[7] if len(node.input) > 7 else ""]) # past_value nis.extend(["past_sequence_length"]) # past_sequence_length nis.extend(["beam_width"]) # beam_width nis.extend(["cache_indirection"]) # cache_indirection nis.extend([node.input[3] if len(node.input) > 3 else ""]) # bias kwargs["past_present_share_buffer"] = 1 node = onnx.helper.make_node( # noqa: PLW2901 "DecoderMaskedMultiHeadAttention", nis, node.output, name=node.name, **kwargs ) if node not in nodes_to_remove: for index, name in enumerate(node.input): if name in tensor_names_to_rename: node.input[index] = target_squeezed_past_seq_name new_nodes.extend([node]) subg.ClearField("node") subg.node.extend(new_nodes) orig_input_names = [inp.name for inp in subg.input] new_inputs = [] for i, vi in enumerate(subg.input): if i >= input_self_past_0 and i < input_cross_past_0: shape = shape_of(vi) vi = onnx.helper.make_tensor_value_info( # noqa: PLW2901 vi.name, elem_type=vi.type.tensor_type.elem_type, shape=[shape[0], shape[1], "max_seq_len", shape[3]], ) new_inputs.extend([vi]) if "past_sequence_length" not in orig_input_names: new_inputs.extend( [onnx.helper.make_tensor_value_info("past_sequence_length", onnx.TensorProto.INT32, shape=[1])] ) if "beam_width" not in orig_input_names: new_inputs.extend([onnx.helper.make_tensor_value_info("beam_width", onnx.TensorProto.INT32, shape=[1])]) if "cache_indirection" not in orig_input_names: new_inputs.extend( [ onnx.helper.make_tensor_value_info( "cache_indirection", onnx.TensorProto.INT32, shape=["batch_size", "beam_width", "max_seq_len"] ) ] ) subg.ClearField("input") subg.input.extend(new_inputs) new_outputs = [] for i, vi in enumerate(subg.output): if i >= output_self_past_0: shape = shape_of(vi) vi = onnx.helper.make_tensor_value_info( # noqa: PLW2901 vi.name, elem_type=vi.type.tensor_type.elem_type, shape=[shape[0], shape[1], "max_seq_len", shape[3]], ) new_outputs.extend([vi]) subg.ClearField("output") subg.output.extend(new_outputs) return True def pack_qkv_for_decoder_masked_mha(model_proto: ModelProto): onnx_model = OnnxModel(model_proto) output_name_to_node = onnx_model.output_name_to_node() nodes_to_add = [] nodes_to_remove = [] for node in onnx_model.nodes(): if node.op_type == "DecoderMaskedMultiHeadAttention": if "past_key_cross" in node.input[1] and "past_value_cross" in node.input[2]: continue q_matmul = output_name_to_node[node.input[0]] k_matmul = output_name_to_node[node.input[1]] v_matmul = output_name_to_node[node.input[2]] q_weight = onnx_model.get_initializer(q_matmul.input[1]) k_weight = onnx_model.get_initializer(k_matmul.input[1]) v_weight = onnx_model.get_initializer(v_matmul.input[1]) if not (q_weight and k_weight and v_weight): return False qw = NumpyHelper.to_array(q_weight) kw = NumpyHelper.to_array(k_weight) vw = NumpyHelper.to_array(v_weight) qkv_weight = np.concatenate([qw, kw, vw], axis=1) matmul_node_name = onnx_model.create_node_name("MatMul", name_prefix="MatMul_QKV") weight = onnx.helper.make_tensor( name=matmul_node_name + "_weight", data_type=TensorProto.FLOAT if q_weight.data_type == 1 else TensorProto.FLOAT16, dims=[qkv_weight.shape[0], qkv_weight.shape[1]], vals=qkv_weight.flatten().tolist(), ) model_proto.graph.initializer.extend([weight]) matmul_node = onnx.helper.make_node( "MatMul", inputs=[q_matmul.input[0], matmul_node_name + "_weight"], outputs=[matmul_node_name + "_out"], name=matmul_node_name, ) node.input[0] = matmul_node.output[0] node.input[1] = "" node.input[2] = "" nodes_to_add.extend([matmul_node]) nodes_to_remove.extend([q_matmul, k_matmul, v_matmul]) onnx_model.add_nodes(nodes_to_add) onnx_model.remove_nodes(nodes_to_remove) onnx_model.update_graph() onnx_model.topological_sort() return True def update_input_shapes_for_gpt2_decoder_model(decoder_onnx_path: str, use_external_data_format: bool = True): """Update the input shapes for the inputs "input_ids" and "position_ids" and make the sequence length dim value 1 for each of them. The decoder model will be over-written. Args: decoder_onnx_path (str): Path of GPT-2 decoder onnx model use_external_data_format(bool): output tensors to external data or not. """ decoder_model_proto = onnx.load_model(decoder_onnx_path, load_external_data=True) for i in range(len(decoder_model_proto.graph.input)): if ( decoder_model_proto.graph.input[i].name == "input_ids" or decoder_model_proto.graph.input[i].name == "position_ids" ): shape_dim_proto = decoder_model_proto.graph.input[i].type.tensor_type.shape.dim[1] # Clear any existing dim_param first if shape_dim_proto.HasField("dim_param"): shape_dim_proto.Clear() # Update dim_value to be 1 shape_dim_proto.dim_value = 1 OnnxModel.save(decoder_model_proto, decoder_onnx_path, save_as_external_data=use_external_data_format) return True def generate_gpt2_init_decoder( decoder_onnx_path: str, init_decoder_onnx_path: str, use_external_data_format: bool = True ) -> bool: """Generates the initial decoder GPT2 subgraph and saves it for downstream use. The initial decoder model will be saved to init_decoder_onnx_path. Args: decoder_onnx_path (str): Path of GPT-2 decoder onnx model init_decoder_onnx_path (str): Path of GPT-2 init decoder onnx model use_external_data_format(bool): output tensors to external data or not. """ init_decoder_model_proto = onnx.load_model(decoder_onnx_path, load_external_data=True) logits_output_name = init_decoder_model_proto.graph.output[0].name gpt2_init_decoder_model = OnnxModel(init_decoder_model_proto) output_name_to_node = gpt2_init_decoder_model.output_name_to_node() assert logits_output_name in output_name_to_node logits_matmul_node = output_name_to_node[logits_output_name] # Sanity check - the logits need to be produced by a MatMul node if logits_matmul_node.op_type != "MatMul": return False # Try to find the last residual Add # For fp16, there are Casts along the way # Normalization Node is : LayerNormalization logits_matmul_to_residual_add_path = gpt2_init_decoder_model.match_parent_path( logits_matmul_node, [ "Cast", "LayerNormalization", "Add", "Add", "Cast", "MatMul", "Cast", "FastGelu", "Cast", "MatMul", "Cast", "LayerNormalization", "Add", ], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0], ) # Normalization Node is : SkipLayerNormalization if logits_matmul_to_residual_add_path is None: logits_matmul_to_residual_add_path = gpt2_init_decoder_model.match_parent_path( logits_matmul_node, [ "Cast", "SkipLayerNormalization", "Cast", "MatMul", "Cast", "FastGelu", "Cast", "MatMul", "Cast", "SkipLayerNormalization", ], [0, 0, 1, 0, 0, 0, 0, 0, 0, 0], ) # Try without the Casts before and after the MatMuls if logits_matmul_to_residual_add_path is None: # Normalization Node is : LayerNormalization logits_matmul_to_residual_add_path = gpt2_init_decoder_model.match_parent_path( logits_matmul_node, ["LayerNormalization", "Add", "Add", "MatMul", "FastGelu", "MatMul", "LayerNormalization", "Add"], [0, 0, 1, 0, 0, 0, 0, 0], ) # Normalization Node is : SkipLayerNormalization if logits_matmul_to_residual_add_path is None: logits_matmul_to_residual_add_path = gpt2_init_decoder_model.match_parent_path( logits_matmul_node, [ "SkipLayerNormalization", "MatMul", "FastGelu", "MatMul", "SkipLayerNormalization", ], [0, 1, 0, 0, 0], ) # TODO(hasesh): Are there more permutations to try before returning ? if logits_matmul_to_residual_add_path is None: return False residual_add_node = logits_matmul_to_residual_add_path[-1] # If the last node in the pattern is SkipLayerNormalization, we need to adjust our pattern searches accordingly is_skiplayernorm_path = residual_add_node.op_type == "SkipLayerNormalization" # Regular LayerNormalization path if not is_skiplayernorm_path: residual_add_to_attention_parent_index = 0 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Add", "Cast", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0, 0] ) # Try other parent index of the residual Add node if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 1 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Add", "Cast", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0, 0], ) # Try without the Casts before and after the MatMuls if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 0 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Add", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0] ) # Try without the Casts before and after the MatMuls and other parent index of the residual Add node if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 1 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Add", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0] ) # SkipLayerNormalization path else: residual_add_to_attention_parent_index = 0 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Cast", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0] ) # Try other parent index of the residual Add node if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 1 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["Cast", "MatMul", "Attention"], [residual_add_to_attention_parent_index, 0, 0] ) # Try without the Casts before and after the MatMuls if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 0 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["MatMul", "Attention"], [residual_add_to_attention_parent_index, 0] ) # Try without the Casts before and after the MatMuls and other parent index of the residual Add node if residual_add_to_attention_path is None: residual_add_to_attention_parent_index = 1 residual_add_to_attention_path = gpt2_init_decoder_model.match_parent_path( residual_add_node, ["MatMul", "Attention"], [residual_add_to_attention_parent_index, 0] ) # TODO(hasesh): Are there more permutations to try before returning ? if residual_add_to_attention_path is None: return False residual_add_to_add_parent_index = 0 if residual_add_to_attention_parent_index == 1 else 1 # Regular LayerNormalization path if not is_skiplayernorm_path: add_before_residual_add = gpt2_init_decoder_model.match_parent( residual_add_node, "Add", residual_add_to_add_parent_index ) # SkipLayerNormalization path else: add_before_residual_add = gpt2_init_decoder_model.match_parent( residual_add_node, "SkipLayerNormalization", residual_add_to_add_parent_index ) if add_before_residual_add is None: return False attention = residual_add_to_attention_path[-1] matmul_after_attention = residual_add_to_attention_path[-2] slice_starts = onnx.helper.make_tensor( name="SliceLastTokenStarts", data_type=TensorProto.INT32, dims=[1], vals=[-1], ) slice_ends = onnx.helper.make_tensor( name="SliceLastTokenEnds", data_type=TensorProto.INT32, dims=[1], vals=[-2], ) slice_axes = onnx.helper.make_tensor( name="SliceLastTokenAxes", data_type=TensorProto.INT32, dims=[1], vals=[1], ) slice_steps = onnx.helper.make_tensor( name="SliceLastTokenSteps", data_type=TensorProto.INT32, dims=[1], vals=[-1], ) gpt2_init_decoder_model.add_initializer(slice_starts) gpt2_init_decoder_model.add_initializer(slice_ends) gpt2_init_decoder_model.add_initializer(slice_axes) gpt2_init_decoder_model.add_initializer(slice_steps) # Add Slice node to the graph such that it consumes the output of Attention slice_0_output_name = "edge_modified_" + attention.output[0] slice_node_0 = onnx.helper.make_node( "Slice", inputs=[ attention.output[0], "SliceLastTokenStarts", "SliceLastTokenEnds", "SliceLastTokenAxes", "SliceLastTokenSteps", ], outputs=[slice_0_output_name], name=gpt2_init_decoder_model.create_node_name("Slice", "GatherLastToken_0_"), ) # Add Slice node to the graph such that it consumes the output of Add before the residual Add # If the 'Add' output is produced by a SkipLayerNormalization node, then adjust its output # index appropriately add_before_residual_add_output = ( add_before_residual_add.output[0] if not is_skiplayernorm_path else add_before_residual_add.output[3] ) slice_1_output_name = "edge_modified_" + add_before_residual_add.output[0] slice_node_1 = onnx.helper.make_node( "Slice", inputs=[ add_before_residual_add_output, "SliceLastTokenStarts", "SliceLastTokenEnds", "SliceLastTokenAxes", "SliceLastTokenSteps", ], outputs=[slice_1_output_name], name=gpt2_init_decoder_model.create_node_name("Slice", "GatherLastToken_1_"), ) # Add the 2 Slice nodes gpt2_init_decoder_model.add_node(slice_node_0) gpt2_init_decoder_model.add_node(slice_node_1) # Adjust the input(s) to the nodes consuming the outputs of the added Slice nodes gpt2_init_decoder_model.replace_node_input(matmul_after_attention, attention.output[0], slice_0_output_name) gpt2_init_decoder_model.replace_node_input(residual_add_node, add_before_residual_add_output, slice_1_output_name) # Topologically sort the updated graph gpt2_init_decoder_model.topological_sort() # Save the init decoder model OnnxModel.save(init_decoder_model_proto, init_decoder_onnx_path, save_as_external_data=use_external_data_format) return True def make_dim_proto_numeric_t5(model, config): """Make dim_proto numeric. Args: model: T5 encoder and decoder model. config: T5 config. """ sequence_length = str(1) num_heads = str(config.num_heads) hidden_size = str(config.d_model) head_size = str(config.d_kv) for tensor in model.graph.output: for dim_proto in tensor.type.tensor_type.shape.dim: if dim_proto.HasField("dim_param") and dim_proto.dim_param in [ sequence_length, num_heads, hidden_size, head_size, ]: dim_value = int(dim_proto.dim_param) dim_proto.Clear() dim_proto.dim_value = dim_value for tensor in model.graph.input: for dim_proto in tensor.type.tensor_type.shape.dim: if dim_proto.HasField("dim_param") and dim_proto.dim_param in [ sequence_length, num_heads, hidden_size, head_size, ]: dim_value = int(dim_proto.dim_param) dim_proto.Clear() dim_proto.dim_value = dim_value def convert_generation_model(args: argparse.Namespace, generation_type: GenerationType = GenerationType.BEAMSEARCH): """Convert model according to command line arguments. Args: args (argparse.Namespace): arguments parsed from command line """ is_gpt2: bool = args.model_type == "gpt2" is_beamsearch: bool = generation_type == GenerationType.BEAMSEARCH is_greedysearch: bool = generation_type == GenerationType.GREEDYSEARCH is_sampling: bool = generation_type == GenerationType.SAMPLING past_present_share_buffer: bool = args.past_present_share_buffer logger.info(f"**** past_present_share_buffer={past_present_share_buffer}") if len(args.op_block_list) == 1 and args.op_block_list[0] == "auto": if is_gpt2 and args.precision == Precision.FLOAT16: args.op_block_list = ["Add", "LayerNormalization", "SkipLayerNormalization", "FastGelu"] logger.info(f"**** Setting op_block_list to {args.op_block_list}") logger.info("**** use --op_block_list if you want to override the block operator list.") else: args.op_block_list = [] if is_greedysearch or is_sampling: if not is_gpt2: raise NotImplementedError("Currently only gpt2 with greedy search/sampling is supported") if args.output_sequences_scores: raise NotImplementedError("output_sequences_scores currently is not supported in greedy search/sampling") if args.output_token_scores: raise NotImplementedError("output_token_scores currently is not supported in greedy search/sampling") # For BeamSearch, sharing buffers for past and present states is only supported # when using `use_decoder_masked_attention` if past_present_share_buffer and is_beamsearch and not args.use_decoder_masked_attention: raise ValueError( "`use_decoder_masked_attention` MUST be turned on to use `past_present_share_buffer` in case of BeamSearch" ) # For any kind of sampling, using decoder masked multihead attention is only supported # when using `past_present_share_buffer` if args.use_decoder_masked_attention and not past_present_share_buffer: raise ValueError("`past_present_share_buffer` MUST be turned on to use `use_decoder_masked_attention`") # For any kind of sampling, using decoder masked multihead attention is only supported # on GPUs if args.use_decoder_masked_attention and not args.use_gpu: raise ValueError("`use_decoder_masked_attention` option is only supported on GPUs") if is_gpt2: if args.decoder_onnx and os.path.exists(args.decoder_onnx): logger.info(f"skip convert_to_onnx since path existed: {args.decoder_onnx}") else: if not args.decoder_onnx: onnx_filename = "{}_past_{}.onnx".format( args.model_name_or_path, "fp16" if args.precision == Precision.FLOAT16 else "fp32" ) args.decoder_onnx = Path(Path(args.output).parent, onnx_filename).as_posix() logger.info(f"Convert GPT model {args.model_name_or_path} to onnx {args.decoder_onnx} ...") gpt2_to_onnx(args) else: # t5 or mt5 if args.decoder_onnx and args.encoder_decoder_init_onnx: logger.info( f"skip convert_to_onnx since paths specified: {args.decoder_onnx} and {args.encoder_decoder_init_onnx}" ) else: logger.info(f"Convert model {args.model_name_or_path} to onnx ...") t5_to_onnx(args) # We only want to pad the logits MatMul weight in the decoder for fp16 models. # The inherent assumption is that fp16 models run on GPU for which all # dims need to be a multiple of 8 to leverage tensor cores. # NOTE: We currently only support padding the MatMul logits weight for GPT2 GreedySearch/BeamSearch. # This can be expanded to other models/decoding strategies later logits_matmul_weight_padded = False if ( not args.disable_pad_vocab_size and args.precision == Precision.FLOAT16 and is_gpt2 and (is_beamsearch or is_greedysearch or is_sampling) ): logger.info( f"Pad logits MatMul weights for optimal MatMul perf in fp16 on {args.decoder_onnx}. " "The file will be overwritten." ) logits_matmul_weight_padded = pad_weights_of_logits_matmul(args.decoder_onnx, args.use_external_data_format) if not logits_matmul_weight_padded: logger.warning( "Tried and failed to pad logits MatMul weights. Performance may be sub-optimal for this MatMul" ) gpt2_init_decoder_generated = False gpt2_init_decoder_onnx_path = None if ( not args.disable_separate_gpt2_decoder_for_init_run and is_gpt2 and (is_beamsearch or is_greedysearch or is_sampling) ): logger.info(f"Creating an initial run GPT2 decoder from {args.decoder_onnx}. ") gpt2_init_decoder_onnx_filename = "gpt2_init_past_{}.onnx".format( "fp16" if args.precision == Precision.FLOAT16 else "fp32" ) gpt2_init_decoder_onnx_path = Path(Path(args.output).parent, gpt2_init_decoder_onnx_filename).as_posix() gpt2_init_decoder_generated = generate_gpt2_init_decoder( args.decoder_onnx, gpt2_init_decoder_onnx_path, args.use_external_data_format ) if not gpt2_init_decoder_generated: logger.warning( "Tried and failed to generate the init decoder GPT2 model. " "Performance may be sub-optimal for the initial decoding run" ) # Update the graph input shapes for the non-initial decoder model to account # for the fact that the sequence length will always be 1 if gpt2_init_decoder_generated and not update_input_shapes_for_gpt2_decoder_model( args.decoder_onnx, args.use_external_data_format ): # Can't proceed further - better to raise an exception raise ValueError("Could not update the input shapes for the non-initial decoder subgraph.") # If the user explicitly requests running shape inference or if we padded/mutated # weight(s)/input shape(s) in the decoder, we want to run shape inference to capture the new # shapes if logits_matmul_weight_padded or args.run_shape_inference or gpt2_init_decoder_generated: logger.info(f"Run symbolic shape inference on {args.decoder_onnx}. The file will be overwritten.") shape_inference(args.decoder_onnx, args.use_external_data_format) if gpt2_init_decoder_generated: logger.info(f"Run symbolic shape inference on {gpt2_init_decoder_onnx_path}. The file will be overwritten.") shape_inference(gpt2_init_decoder_onnx_path, args.use_external_data_format) if is_gpt2: config = GPT2Config.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir) elif args.model_type == "t5": config = T5Config.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir) else: config = MT5Config.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir) if args.verbose: logger.info(f"Config={config}") eos_token_id = config.eos_token_id pad_token_id = config.eos_token_id if is_gpt2 else config.pad_token_id vocab_size = config.vocab_size # if vocab_size is given in parameters use that. if args.vocab_size != -1: vocab_size = args.vocab_size if args.eos_token_id != -1: eos_token_id = args.eos_token_id if args.pad_token_id != -1: pad_token_id = args.pad_token_id decoder_model = onnx.load_model(args.decoder_onnx, load_external_data=True) decoder_model.graph.name = f"{args.model_type} decoder" gpt2_init_decoder_model = None if args.model_type == "gpt2": verify_gpt2_subgraph(decoder_model.graph, args.precision) # If we generated the init decoder model, verify that as well if gpt2_init_decoder_generated: gpt2_init_decoder_model = onnx.load_model(gpt2_init_decoder_onnx_path, load_external_data=True) gpt2_init_decoder_model.graph.name = f"{args.model_type} init decoder" verify_gpt2_subgraph(gpt2_init_decoder_model.graph, args.precision) else: verify_t5_decoder_subgraph(decoder_model.graph, args.precision) inputs = None if is_beamsearch: inputs = [ "input_ids", "max_length", "min_length", "num_beams", "num_return_sequences", "length_penalty", "repetition_penalty", ] elif is_greedysearch or is_sampling: inputs = [ "input_ids", "max_length", "min_length", "repetition_penalty", ] if args.vocab_mask: inputs.append("vocab_mask") else: inputs.append("") if args.prefix_vocab_mask: inputs.append("prefix_vocab_mask") else: inputs.append("") if args.custom_attention_mask: inputs.append("attention_mask") else: inputs.append("") if is_sampling: if args.custom and args.presence_mask: inputs.append("presence_mask") else: inputs.append("") if args.seed: inputs.append("seed") outputs = ["sequences"] if args.output_sequences_scores: outputs.append("sequences_scores") if args.output_token_scores: assert args.output_sequences_scores, "--output_token_scores requires --output_sequences_scores" outputs.append("scores") node = None if is_beamsearch: node = onnx.helper.make_node( "BeamSearch", inputs=inputs, outputs=outputs, name=f"BeamSearch_{args.model_type}", ) elif is_greedysearch: node = onnx.helper.make_node( "GreedySearch", inputs=inputs, outputs=outputs, name=f"GreedySearch_{args.model_type}", ) elif is_sampling: node = onnx.helper.make_node( "Sampling", inputs=inputs, outputs=outputs, name=f"Sampling_{args.model_type}", ) node.domain = "com.microsoft" attr_to_extend = None if is_beamsearch: attr_to_extend = [ onnx.helper.make_attribute("eos_token_id", eos_token_id), onnx.helper.make_attribute("pad_token_id", pad_token_id), onnx.helper.make_attribute("no_repeat_ngram_size", args.no_repeat_ngram_size), onnx.helper.make_attribute("early_stopping", 1 if args.early_stopping else 0), onnx.helper.make_attribute("model_type", 0 if args.model_type == "gpt2" else 1), ] elif is_greedysearch: attr_to_extend = [ onnx.helper.make_attribute("eos_token_id", eos_token_id), onnx.helper.make_attribute("pad_token_id", pad_token_id), onnx.helper.make_attribute("model_type", 0 if args.model_type == "gpt2" else 1), onnx.helper.make_attribute("no_repeat_ngram_size", args.no_repeat_ngram_size), ] elif is_sampling: attr_to_extend = [ onnx.helper.make_attribute("eos_token_id", eos_token_id), onnx.helper.make_attribute("pad_token_id", pad_token_id), onnx.helper.make_attribute("model_type", 0 if args.model_type == "gpt2" else 1), onnx.helper.make_attribute("no_repeat_ngram_size", args.no_repeat_ngram_size), onnx.helper.make_attribute("temperature", args.temperature), onnx.helper.make_attribute("top_p", args.top_p), onnx.helper.make_attribute("filter_value", args.filter_value), onnx.helper.make_attribute("min_tokens_to_keep", args.min_tokens_to_keep), onnx.helper.make_attribute("custom", args.custom), onnx.helper.make_attribute("presence_penalty", args.presence_penalty), ] # Explicitly pass in the vocab size via an attribute if logits_matmul_weight_padded: attr_to_extend.extend([onnx.helper.make_attribute("vocab_size", vocab_size)]) node.attribute.extend(attr_to_extend) initializers = [] if args.model_type in ["t5", "mt5"]: if args.run_shape_inference: logger.info(f"Symbolic shape inference on {args.encoder_decoder_init_onnx}. The file will be overwritten.") shape_inference(args.encoder_decoder_init_onnx, args.use_external_data_format) encoder_model = onnx.load_model(args.encoder_decoder_init_onnx, load_external_data=True) encoder_model.graph.name = f"{args.model_type} encoder and decoder init" verify_t5_encoder_decoder_init_subgraph(encoder_model.graph, args.precision) make_dim_proto_numeric_t5(encoder_model, config) make_dim_proto_numeric_t5(decoder_model, config) # Update decoder subgraph in preparation to use past present share buffer if past_present_share_buffer: if not args.use_decoder_masked_attention: raise ValueError("past_present_share_buffer is only supported with use_decoder_masked_attention") logger.info( "*****update t5 decoder subgraph to share past/present buffer and use decoder_masked_multihead_attention*****" ) if update_decoder_subgraph_share_buffer_and_use_decoder_masked_mha(decoder_model.graph): logger.info("*****update t5 decoder subgraph successfully!!!*****") else: logger.info("*****DecoderMaskedMultiHeadAttention is not applied to T5 decoder*****") if pack_qkv_for_decoder_masked_mha(decoder_model): logger.info("*****pack qkv for decoder masked mha successfully!!!*****") else: logger.info("*****pack qkv for decoder masked mha failed!!!*****") if not args.disable_shared_initializers: # Unique shared initializers from the decoder and decoder_init could reduce memory usage in inference. initializers = get_shared_initializers(encoder_model, decoder_model) logger.info( f"{len(initializers)} shared initializers ({[i.name for i in initializers]}) in encoder and decoder subgraphs are moved to the main graph" ) # TODO(tianleiwu): investigate the following which causes error in inference # Move initializer from subgraph to main graph could reduce memory usage in inference. # moved_initializers = move_initializers(encoder_model.graph) # logger.info( # f"{len(moved_initializers)} initializers ({[i.name for i in moved_initializers]}) from the encoder are moved to the main graph" # ) # initializers.extend(moved_initializers) node.attribute.extend( [ onnx.helper.make_attribute("encoder", encoder_model.graph), onnx.helper.make_attribute("decoder", decoder_model.graph), onnx.helper.make_attribute( "decoder_start_token_id", config.decoder_start_token_id if len(encoder_model.graph.input) == 3 else -1, ), ] ) else: if gpt2_init_decoder_generated: # Move shared initializers (shared between init decoder and decoder models) to the main # graph and remove them from these models if not args.disable_shared_initializers: # Unique shared initializers from the decoder and decoder_init could reduce memory usage in inference. initializers = get_shared_initializers(gpt2_init_decoder_model, decoder_model) logger.info( f"{len(initializers)} shared initializers ({[i.name for i in initializers]}) in decoder and init decoder subgraphs are moved to the main graph" ) # Update init decoder subgraph in preparation to use past present share buffer if past_present_share_buffer: logger.info("*****update init decoder subgraph to make past and present share buffer******************") update_decoder_subgraph_past_present_share_buffer(gpt2_init_decoder_model.graph) # Update init decoder subgraph in preparation to use DecoderMaskedSelfAttention # NOTE: Even if we will not use DecoderMaskedSelfAttention in the init decoder subgraph # it makes the runtime changes cleaner if we keep both the init decoder and decoder subgraphs # same in terms of the subgraph inputs. if args.use_decoder_masked_attention and not update_decoder_subgraph_use_decoder_masked_attention( gpt2_init_decoder_model.graph, is_beamsearch, False ): raise ValueError("Could not update the init decoder subgraph to use DecoderMaskedSelfAttention") node.attribute.append(onnx.helper.make_attribute("init_decoder", gpt2_init_decoder_model.graph)) else: # Move initializer from subgraph to main graph could reduce memory usage in inference. initializers = move_initializers(decoder_model.graph) logger.info(f"{len(initializers)} initializers from the decoder are moved to the main graph") # Update decoder subgraph in preparation to use past present share buffer if past_present_share_buffer: logger.info("*****update decoder subgraph to make past and present share buffer******************") update_decoder_subgraph_past_present_share_buffer(decoder_model.graph) # Update decoder subgraph in preparation to use DecoderMaskedSelfAttention if args.use_decoder_masked_attention and not update_decoder_subgraph_use_decoder_masked_attention( decoder_model.graph, is_beamsearch, True ): raise ValueError("Could not update the decoder subgraph to use DecoderMaskedSelfAttention") node.attribute.append(onnx.helper.make_attribute("decoder", decoder_model.graph)) # graph inputs input_ids = onnx.helper.make_tensor_value_info("input_ids", TensorProto.INT32, ["batch_size", "sequence_length"]) max_length = onnx.helper.make_tensor_value_info("max_length", TensorProto.INT32, [1]) min_length = onnx.helper.make_tensor_value_info("min_length", TensorProto.INT32, [1]) num_beams = onnx.helper.make_tensor_value_info("num_beams", TensorProto.INT32, [1]) num_return_sequences = onnx.helper.make_tensor_value_info("num_return_sequences", TensorProto.INT32, [1]) length_penalty = onnx.helper.make_tensor_value_info("length_penalty", TensorProto.FLOAT, [1]) repetition_penalty = onnx.helper.make_tensor_value_info("repetition_penalty", TensorProto.FLOAT, [1]) graph_inputs = None if is_beamsearch: graph_inputs = [ input_ids, max_length, min_length, num_beams, num_return_sequences, length_penalty, repetition_penalty, ] elif is_greedysearch or is_sampling: graph_inputs = [ input_ids, max_length, min_length, repetition_penalty, ] if args.vocab_mask: vocab_mask = onnx.helper.make_tensor_value_info("vocab_mask", TensorProto.INT32, [vocab_size]) graph_inputs.append(vocab_mask) if args.prefix_vocab_mask: prefix_vocab_mask = onnx.helper.make_tensor_value_info( "prefix_vocab_mask", TensorProto.INT32, ["batch_size", vocab_size] ) graph_inputs.append(prefix_vocab_mask) if args.custom_attention_mask: attention_mask = onnx.helper.make_tensor_value_info( "attention_mask", TensorProto.INT32, ["batch_size", "sequence_length"] ) graph_inputs.append(attention_mask) if args.custom and args.presence_mask: presence_mask = onnx.helper.make_tensor_value_info( "presence_mask", TensorProto.INT32, ["batch_size", vocab_size] ) graph_inputs.append(presence_mask) if is_sampling and args.seed: seed = onnx.helper.make_tensor_value_info("seed", TensorProto.INT32, [1]) graph_inputs.append(seed) # graph outputs sequences = None if is_beamsearch: sequences = onnx.helper.make_tensor_value_info( "sequences", TensorProto.INT32, ["batch_size", "num_return_sequences", "max_length"], ) elif is_greedysearch or is_sampling: sequences = onnx.helper.make_tensor_value_info( "sequences", TensorProto.INT32, ["batch_size", "max_length"], ) graph_outputs = [sequences] if args.output_sequences_scores: sequences_scores = onnx.helper.make_tensor_value_info( "sequences_scores", TensorProto.FLOAT, ["batch_size", "num_return_sequences"] ) graph_outputs.append(sequences_scores) if args.output_token_scores: scores = onnx.helper.make_tensor_value_info( "scores", TensorProto.FLOAT, ["max_length - sequence_length", "batch_size", "num_beams", vocab_size], ) graph_outputs.append(scores) new_graph = onnx.helper.make_graph( [node], f"{args.model_type} beam search" if not is_greedysearch else f"{args.model_type} greedy search", graph_inputs, graph_outputs, initializers, ) # Create the model new_model = onnx.helper.make_model( new_graph, producer_name="onnxruntime.transformers", opset_imports=decoder_model.opset_import, ) # TODO(tianleiwu): move shared initializers from T5 encoder and decoder subgraphs to parent graph to save memory. if args.use_external_data_format: from packaging import version if version.parse(onnx.__version__) < version.parse("1.12.0"): logger.warning("Require onnx >= 1.12 to save large (>2GB) model!") OnnxModel.save( new_model, args.output, save_as_external_data=True, all_tensors_to_one_file=True, ) else: onnx.save(new_model, args.output) logger.info(f"model save to {args.output}") def test_torch_performance( args: argparse.Namespace, model: Union[GPT2LMHeadModel, T5ForConditionalGeneration], input_ids: torch.Tensor, attention_mask: torch.Tensor, eos_token_id: int, pad_token_id: int, bad_words_ids: List[List[int]], ) -> Dict[str, Any]: """Test PyTorch performance of text generation. Args: args (argparse.Namespace): arguments parsed from command line model (Union[GPT2LMHeadModel, T5ForConditionalGeneration]): PyTorch model input_ids (torch.Tensor): input_ids attention_mask (torch.Tensor): Attention mask eos_token_id (int): EOS token ID pad_token_id (int): Padding token ID bad_words_ids (List[List[int]]): Words shall not be generated. Raises: RuntimeError: PyTorch with CUDA is not available for --use_gpu Returns: Dict[str, Any]: A dictionary with string with metric name, and value can be integer or string. """ if args.use_gpu and not torch.cuda.is_available(): raise RuntimeError("Please install PyTorch with Cuda for testing gpu performance.") if args.precision == Precision.FLOAT16: model.half() device = torch.device("cuda:0" if args.use_gpu else "cpu") model.to(device) torch.set_grad_enabled(False) input_ids = input_ids.to(device) attention_mask = attention_mask.to(device) torch_latency = [] for _ in range(args.total_runs): start = time.time() _ = model.generate( input_ids=input_ids, attention_mask=attention_mask, max_length=args.max_length, min_length=args.min_length, num_beams=args.num_beams, early_stopping=args.early_stopping, no_repeat_ngram_size=args.no_repeat_ngram_size, eos_token_id=eos_token_id, pad_token_id=pad_token_id, num_return_sequences=args.num_return_sequences, length_penalty=args.length_penalty, repetition_penalty=args.repetition_penalty, bad_words_ids=bad_words_ids if bad_words_ids else None, return_dict_in_generate=True, output_scores=args.output_sequences_scores or args.output_token_scores, ) torch_latency.append(time.time() - start) batch_size = input_ids.shape[0] from benchmark_helper import get_latency_result return get_latency_result(torch_latency, batch_size) def create_attention_mask(input_ids, pad_token_id): attention_mask = np.ones(input_ids.shape, dtype=np.int32) for i in range(input_ids.shape[0]): abs_pos = 0 for j in range(input_ids.shape[1]): if input_ids[i][j] == pad_token_id and abs_pos == 0: attention_mask[i][j] = 0 else: abs_pos += 1 return attention_mask def test_gpt_model(args: argparse.Namespace, sentences: Optional[List[str]] = None, is_greedy: bool = False): """Test GPT-2 model Args: args (argparse.Namespace): arguments parsed from command line sentences (Optional[List[str]], optional): input text. Defaults to None. Returns: Union[Dict[str, Any], None]: A dictionary with string with metric name, and value can be integer or string. """ assert args.model_type == "gpt2" tokenizer = GPT2Tokenizer.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir) tokenizer.padding_side = "left" tokenizer.pad_token = tokenizer.eos_token model = GPT2LMHeadModel.from_pretrained( args.model_name_or_path, cache_dir=args.cache_dir, pad_token_id=tokenizer.eos_token_id, ) # Use different length sentences to test batching if sentences is None: sentences = [ "The product is released", "I enjoy walking in the park", "Test best way to invest", ] inputs = tokenizer(sentences, return_tensors="pt", padding=True) input_ids = inputs["input_ids"] attention_mask = inputs["attention_mask"] bad_words = "walk in park" bad_words_ids = tokenizer.encode(bad_words, add_prefix_space=True) bad_words_ids = [[word_id] for word_id in bad_words_ids] # Convert to list of list if args.vocab_mask: logger.debug("bad_words_ids", bad_words_ids) # noqa: PLE1205 else: bad_words_ids = [] config = model.config eos_token_id = config.eos_token_id pad_token_id = config.eos_token_id vocab_size = config.vocab_size torch_decoded_sequences = [] beam_outputs = None if not args.disable_parity: print("-" * 50) print("Test PyTorch model and beam search with huggingface transformers...") beam_outputs = model.generate( input_ids=input_ids, attention_mask=attention_mask, max_length=args.max_length, min_length=args.min_length, num_beams=args.num_beams, early_stopping=args.early_stopping, no_repeat_ngram_size=args.no_repeat_ngram_size, eos_token_id=eos_token_id, pad_token_id=pad_token_id, num_return_sequences=args.num_return_sequences, length_penalty=args.length_penalty, repetition_penalty=args.repetition_penalty, bad_words_ids=bad_words_ids if bad_words_ids else None, return_dict_in_generate=True, output_scores=args.output_sequences_scores or args.output_token_scores, ) print("input_ids", input_ids) print("huggingface transformers outputs:") print("sequences", beam_outputs.sequences) if args.output_sequences_scores: print("sequences_scores", beam_outputs.sequences_scores) if args.output_token_scores: print("scores", beam_outputs.scores) for i, sequence in enumerate(beam_outputs.sequences): decoded_sequence = tokenizer.decode(sequence, skip_special_tokens=True) torch_decoded_sequences.append(decoded_sequence) print(f"{i}: {decoded_sequence}") print("-" * 50) print("Testing beam search with onnxruntime...") if is_greedy: inputs = { "input_ids": input_ids.cpu().numpy().astype(np.int32), "max_length": np.array([args.max_length], dtype=np.int32), "min_length": np.array([args.min_length], dtype=np.int32), "repetition_penalty": np.array([args.repetition_penalty], dtype=np.float32), } else: inputs = { "input_ids": input_ids.cpu().numpy().astype(np.int32), "max_length": np.array([args.max_length], dtype=np.int32), "min_length": np.array([args.min_length], dtype=np.int32), "num_beams": np.array([args.num_beams], dtype=np.int32), "num_return_sequences": np.array([args.num_return_sequences], dtype=np.int32), "length_penalty": np.array([args.length_penalty], dtype=np.float32), "repetition_penalty": np.array([args.repetition_penalty], dtype=np.float32), } if args.vocab_mask: vocab_mask = np.ones((vocab_size), dtype=np.int32) if args.vocab_mask: for bad_word_id in bad_words_ids: vocab_mask[bad_word_id] = 0 inputs["vocab_mask"] = vocab_mask if args.custom_attention_mask: inputs["attention_mask"] = create_attention_mask(input_ids, pad_token_id) batch_size = input_ids.shape[0] if args.prefix_vocab_mask: logger.info("Use prefix vocab mask with all ones in ORT, but no corresponding setting for Torch model.") prefix_vocab_mask = np.ones((batch_size, vocab_size), dtype=np.int32) inputs["prefix_vocab_mask"] = prefix_vocab_mask if args.save_test_data: test_data_dir = Path(args.output).parent.as_posix() logger.debug("test_data_dir", test_data_dir) # noqa: PLE1205 from bert_test_data import output_test_data logger.info(f"Saving test_data to {test_data_dir}/test_data_set_* ...") all_inputs = [inputs] for i, inputs in enumerate(all_inputs): dir = os.path.join(test_data_dir, "test_data_set_" + str(i)) output_test_data(dir, inputs) logger.debug("ORT inputs", inputs) # noqa: PLE1205 if args.disable_perf_test: return logger.debug("Creating ort session......") ort_session = create_ort_session(args.output, args.use_gpu, args.use_sln_strict_mode) logger.debug("Run ort session......") result = ort_session.run(None, inputs) # Test performance latency = [] for _ in range(args.total_runs): start = time.time() _ = ort_session.run(None, inputs) latency.append(time.time() - start) from benchmark_helper import get_latency_result batch_size = input_ids.shape[0] output = get_latency_result(latency, batch_size) print("ORT outputs:") sequences = result[0] print("sequences", sequences) if args.output_sequences_scores: print("sequences_scores", result[1]) if args.output_token_scores: print("scores", result[2]) if is_greedy: (batch_size, max_length) = sequences.shape ort_decoded_sequences = [] for i in range(batch_size): decoded_sequence = tokenizer.decode(sequences[i], skip_special_tokens=True) ort_decoded_sequences.append(decoded_sequence) print(f"batch {i} sequence: {decoded_sequence}") else: (batch_size, num_sequences, max_length) = sequences.shape ort_decoded_sequences = [] for i in range(batch_size): for j in range(num_sequences): decoded_sequence = tokenizer.decode(sequences[i][j], skip_special_tokens=True) ort_decoded_sequences.append(decoded_sequence) print(f"batch {i} sequence {j}: {decoded_sequence}") if beam_outputs: torch_sequences = beam_outputs.sequences.reshape(batch_size, args.num_return_sequences, -1) ort_sequences = torch.LongTensor(sequences) print("-" * 50) print("Torch Sequences:") print(torch_sequences) print(torch_decoded_sequences) print("-" * 50) print("ORT Sequences:") print(ort_sequences) print(ort_decoded_sequences) print("-" * 50) # Compare the generated text instead of word IDs since ORT pads to max sequence length but Torch not. is_same = torch_decoded_sequences == ort_decoded_sequences print("Torch and ORT result is ", "same" if is_same else "different") output["parity"] = is_same if args.torch_performance: torch_latency_output = test_torch_performance( args, model, input_ids, attention_mask, eos_token_id, pad_token_id, bad_words_ids, ) print("Torch Latency", torch_latency_output) print("ORT", output) return output def test_t5_model(args: argparse.Namespace, sentences: Optional[List[str]] = None): """Test T5 or MT5 model Args: args (argparse.Namespace): arguments parsed from command line sentences (Optional[List[str]], optional): input text. Defaults to None. Returns: Union[Dict[str, Any], None]: A dictionary with string with metric name, and value can be integer or string. """ assert args.model_type in ["t5", "mt5"] if args.prefix_vocab_mask: logger.debug("Skipping parity test as prefix vocab mask is not implemented by Hugging Face") return None tokenizer = T5Tokenizer.from_pretrained(args.model_name_or_path, cache_dir=args.cache_dir) tokenizer.padding_side = "left" if args.model_type == "t5": model = T5ForConditionalGeneration.from_pretrained( args.model_name_or_path, cache_dir=args.cache_dir, ) else: model = MT5ForConditionalGeneration.from_pretrained( args.model_name_or_path, cache_dir=args.cache_dir, ) # Use different length sentences to test batching if sentences is None: sentences = [ "translate English to French: The product is released", "summarize: research continues to show that pets bring real health benefits to their owners. Having a dog around can lead to lower levels of stress for both adults and kids.", # "summarize: I enjoy walking in the park. It makes my mind feel calm and refreshed. " # + "I enjoy looking at the trees, flowers, and wildlife around me, and listening to sound from natural.", ] inputs = tokenizer(sentences, return_tensors="pt", padding=True) input_ids = inputs["input_ids"] attention_mask = inputs["attention_mask"] bad_words = "walk in park" bad_words_ids = tokenizer.encode(bad_words)[:-1] # exclude the last token (EOS) bad_words_ids = [[word_id] for word_id in bad_words_ids] # Convert to list of list if args.vocab_mask: logger.debug("bad_words_ids", bad_words_ids) # noqa: PLE1205 else: bad_words_ids = [] config = model.config eos_token_id = config.eos_token_id pad_token_id = config.pad_token_id vocab_size = config.vocab_size logger.debug(f"eos_token_id:{eos_token_id}, pad_token_id:{pad_token_id}, vocab_size:{vocab_size}") torch_decoded_sequences = [] if not args.disable_parity: print("-" * 50) print("Test PyTorch model and beam search with huggingface transformers...") beam_outputs = model.generate( input_ids=input_ids, attention_mask=attention_mask, max_length=args.max_length, min_length=args.min_length, num_beams=args.num_beams, early_stopping=args.early_stopping, no_repeat_ngram_size=args.no_repeat_ngram_size, eos_token_id=eos_token_id, pad_token_id=pad_token_id, num_return_sequences=args.num_return_sequences, length_penalty=args.length_penalty, repetition_penalty=args.repetition_penalty, bad_words_ids=bad_words_ids if bad_words_ids else None, return_dict_in_generate=True, output_scores=args.output_sequences_scores or args.output_token_scores, ) print("input_ids", input_ids) print("huggingface transformers outputs:") print("sequences", beam_outputs.sequences) if args.output_sequences_scores: print("sequences_scores", beam_outputs.sequences_scores) if args.output_token_scores: print("scores", beam_outputs.scores) for i, sequence in enumerate(beam_outputs.sequences): decoded_sequence = tokenizer.decode(sequence, skip_special_tokens=True) torch_decoded_sequences.append(decoded_sequence) print(f"{i}: {decoded_sequence}") print("-" * 50) print("Testing beam search with onnxruntime...") vocab_mask = np.ones((vocab_size), dtype=np.int32) if args.vocab_mask: for bad_word_id in bad_words_ids: vocab_mask[bad_word_id] = 0 inputs = { "input_ids": input_ids.cpu().numpy().astype(np.int32), "max_length": np.array([args.max_length], dtype=np.int32), "min_length": np.array([args.min_length], dtype=np.int32), "num_beams": np.array([args.num_beams], dtype=np.int32), "num_return_sequences": np.array([args.num_return_sequences], dtype=np.int32), "length_penalty": np.array([args.length_penalty], dtype=np.float32), "repetition_penalty": np.array([args.repetition_penalty], dtype=np.float32), } if args.vocab_mask: inputs["vocab_mask"] = vocab_mask if args.custom_attention_mask: inputs["attention_mask"] = create_attention_mask(input_ids, pad_token_id) if args.save_test_data: test_data_dir = Path(args.output).parent.as_posix() logger.debug("test_data_dir", test_data_dir) # noqa: PLE1205 from bert_test_data import output_test_data all_inputs = [inputs] for i, inputs in enumerate(all_inputs): dir = os.path.join(test_data_dir, "test_data_set_" + str(i)) output_test_data(dir, inputs) logger.debug("ORT inputs", inputs) # noqa: PLE1205 ort_session = create_ort_session(args.output, args.use_gpu, args.use_sln_strict_mode) # Test performance latency = [] for _ in range(args.total_runs): start = time.time() result = ort_session.run(None, inputs) latency.append(time.time() - start) batch_size = input_ids.shape[0] from benchmark_helper import get_latency_result output = get_latency_result(latency, batch_size) print("ORT outputs:") sequences = result[0] print("sequences", sequences) if args.output_sequences_scores: print("sequences_scores", result[1]) if args.output_token_scores: print("scores", result[2]) (batch_size, num_sequences, max_length) = sequences.shape ort_decoded_sequences = [] for i in range(batch_size): for j in range(num_sequences): decoded_sequence = tokenizer.decode(sequences[i][j], skip_special_tokens=True) ort_decoded_sequences.append(decoded_sequence) print(f"batch {i} sequence {j}: {decoded_sequence}") if not args.disable_parity: torch_sequences = beam_outputs.sequences.reshape(batch_size, args.num_return_sequences, -1) ort_sequences = torch.LongTensor(sequences) print("-" * 50) print("Torch Sequences:") print(torch_sequences) print(torch_decoded_sequences) print("-" * 50) print("ORT Sequences:") print(ort_sequences) print(ort_decoded_sequences) print("-" * 50) # Compare the generated text instead of word IDs since ORT pads to max sequence length but Torch not. is_same = torch_decoded_sequences == ort_decoded_sequences print("Torch and ORT result is ", "same" if is_same else "different") output["parity"] = is_same if args.torch_performance: torch_latency_output = test_torch_performance( args, model, input_ids, attention_mask, eos_token_id, pad_token_id, bad_words_ids, ) print("Torch Latency", torch_latency_output) print("ORT", output) return output def main(argv: Optional[List[str]] = None, sentences: Optional[List[str]] = None): """Main entry function Args: argv (Optional[List[str]], optional): _description_. Defaults to None. sentences (Optional[List[str]], optional): input text. Defaults to None. Raises: ValueError: Path does not exist: --encoder_decoder_init_onnx ValueError: Path does not exist: --decoder_onnx ValueError: --decoder_onnx and --encoder_decoder_init_onnx are not used together for T5 Returns: Union[Dict[str, Any], None]: A dictionary with string with metric name, and value can be integer or string. """ args = parse_arguments(argv) setup_logger(args.verbose) if args.model_type in ["t5", "mt5"]: if args.encoder_decoder_init_onnx and not os.path.exists(args.encoder_decoder_init_onnx): raise ValueError(f"Path does not exist: --encoder_decoder_init_onnx {args.encoder_decoder_init_onnx}") if args.decoder_onnx and not os.path.exists(args.decoder_onnx): raise ValueError(f"Path does not exist: --decoder_onnx {args.decoder_onnx}") if (args.encoder_decoder_init_onnx and not args.decoder_onnx) or ( args.decoder_onnx and not args.encoder_decoder_init_onnx ): raise ValueError("--decoder_onnx shall use together with --encoder_decoder_init_onnx") is_greedy = args.num_beams == 1 and args.num_return_sequences == 1 if args.model_type == "gpt2" and is_greedy: if args.top_p > 0.0 and args.top_p < 1.0: convert_generation_model(args, GenerationType.SAMPLING) logger.info( "The test for gpt2_sampling onnx model is limited to non-custom model with small top_p(e.g <=0.01) value. The result should be the same as gpt2 greedy search." ) if args.top_p > 0.01 or args.custom or args.seed: return else: convert_generation_model(args, GenerationType.GREEDYSEARCH) else: convert_generation_model(args) logger.info("start testing model...") if args.model_type in ["t5", "mt5"]: result = test_t5_model(args, sentences=sentences) else: result = test_gpt_model(args, sentences=sentences, is_greedy=is_greedy) if result: if args.use_external_data_format: logger.info(f"Output files: {args.output}, {args.output}.data") else: logger.info(f"Output file: {args.output}") return result if __name__ == "__main__": main()