Update bert benchmark: replace deprecated API (#22611)

tianleiwu · web-flow · commit 975d3dffcff9 · 2024-10-28T13:24:17.000-07:00
### Description
(1) tokenizer.max_model_input_sizes was deprecated. Use
tokenizer.model_max_length to replace it.
(2) onnx opset updated to 16 instead of 11/12 for models.
(3) Update a few comments related to torch installation.
(4) Test gpu instead of cpu in dev_benchmark.cmd.

### Motivation and Context
Update bert benchmark script so that it can run with latest huggingface
transformers package.
diff --git a/onnxruntime/python/tools/transformers/benchmark.py b/onnxruntime/python/tools/transformers/benchmark.py
@@ -348,7 +348,7 @@ def run_pytorch(
         else:
             tokenizer = AutoTokenizer.from_pretrained(model_name, cache_dir=cache_dir)
 
-            max_input_size = tokenizer.max_model_input_sizes.get(model_name, 1024)
+            max_input_size = tokenizer.model_max_length
 
         logger.debug(f"Model {model}")
         logger.debug(f"Number of parameters {model.num_parameters()}")
@@ -500,7 +500,7 @@ def run_tensorflow(
 
         tokenizer = AutoTokenizer.from_pretrained(model_name, cache_dir=cache_dir)
 
-        max_input_size = tokenizer.max_model_input_sizes.get(model_name, 1024)
+        max_input_size = tokenizer.model_max_length
 
         for batch_size in batch_sizes:
             if batch_size <= 0:
diff --git a/onnxruntime/python/tools/transformers/dev_benchmark.cmd b/onnxruntime/python/tools/transformers/dev_benchmark.cmd
@@ -3,9 +3,7 @@
 REM Run benchmark in Windows for developing purpose. For official benchmark, please use run_benchmark.sh.
 REM Settings are different from run_benchmark.sh: no cli, batch and sequence, input counts, average over 100, no fp16, less models etc.
 
-REM Please install PyTorch (see https://pytorch.org/) before running this benchmark. Like the following:
-REM   GPU:   conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
-REM   CPU:   conda install pytorch torchvision cpuonly -c pytorch
+REM Please install PyTorch (see https://pytorch.org/) before running this benchmark.
 
 REM When use_package=true, you need not copy other files to run benchmarks except this sh file.
 REM Otherwise, it will use python script (*.py) files in this directory.
@@ -21,12 +19,12 @@ set run_torchscript=false
 
 REM Devices to test.
 REM Attention: You cannot run both CPU and GPU at the same time: gpu need onnxruntime-gpu, and CPU need onnxruntime.
-set run_gpu_fp32=false
-set run_gpu_fp16=false
-set run_cpu_fp32=true
-set run_cpu_int8=true
+set run_gpu_fp32=true
+set run_gpu_fp16=true
+set run_cpu_fp32=false
+set run_cpu_int8=false
 
-set average_over=100
+set average_over=1000
 
 REM Enable optimizer (use script instead of OnnxRuntime for graph optimization)
 set use_optimizer=true
@@ -36,7 +34,7 @@ set sequence_length=8 128
 
 REM Number of inputs (input_ids, token_type_ids, attention_mask) for ONNX model.
 REM Note that different input count might lead to different performance
-set input_counts=1
+set input_counts=3
 
 REM Pretrained transformers models can be a subset of: bert-base-cased roberta-base gpt2 distilgpt2 distilbert-base-uncased
 set models_to_test=bert-base-cased
@@ -57,7 +55,6 @@ if %run_cpu_int8% == true if %run_gpu_fp32% == true echo cannot test cpu and gpu
 if %run_cpu_int8% == true if %run_gpu_fp16% == true echo cannot test cpu and gpu at same time & goto :EOF
 
 if %run_install% == true (
-  pip uninstall --yes ort_nightly
   pip uninstall --yes onnxruntime
   pip uninstall --yes onnxruntime-gpu
   if %run_cpu_fp32% == true (
@@ -70,7 +67,6 @@ if %run_install% == true (
     )
   )
 
-  pip install --upgrade onnxconverter_common
   pip install --upgrade transformers
 )
 
diff --git a/onnxruntime/python/tools/transformers/huggingface_models.py b/onnxruntime/python/tools/transformers/huggingface_models.py
@@ -13,155 +13,62 @@
     "AutoModelForCausalLM",
 ]
 
-# List of pretrained models: https://huggingface.co/transformers/pretrained_models.html
 # Pretrained model name to a tuple of input names, opset_version, use_external_data_format, optimization model type
+# Some models like GPT, T5, Bart etc has its own convert_to_onnx.py in models sub-directory, and they are excluded here.
 MODELS = {
     # BERT
-    "bert-base-uncased": (
-        ["input_ids", "attention_mask", "token_type_ids"],
-        12,
-        False,
-        "bert",
-    ),
-    "bert-large-uncased": (
-        ["input_ids", "attention_mask", "token_type_ids"],
-        12,
-        False,
-        "bert",
-    ),
-    "bert-base-cased": (
-        ["input_ids", "attention_mask", "token_type_ids"],
-        12,
-        False,
-        "bert",
-    ),
-    # "bert-large-cased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-multilingual-uncased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-multilingual-cased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-chinese": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-german-cased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-large-uncased-whole-word-masking": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-large-cased-whole-word-masking": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-large-uncased-whole-word-masking-finetuned-squad": (["input_ids", "attention_mask",
-    #                                                            "token_type_ids"], 12, False, "bert"),
-    # "bert-large-cased-whole-word-masking-finetuned-squad": (["input_ids", "attention_mask",
-    #                                                          "token_type_ids"], 12, False, "bert"),
-    # "bert-base-cased-finetuned-mrpc": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-german-dbmdz-cased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # "bert-base-german-dbmdz-uncased": (["input_ids", "attention_mask", "token_type_ids"], 12, False, "bert"),
-    # todo: more models to add
-    # GPT (no past state)
-    "openai-gpt": (["input_ids"], 11, False, "gpt2"),
-    # GPT-2 (no past state, use benchmark_gpt2.py for past_key_values)
-    "gpt2": (["input_ids"], 11, False, "gpt2"),
-    "gpt2-medium": (["input_ids"], 11, False, "gpt2"),
-    "gpt2-large": (["input_ids"], 11, True, "gpt2"),
-    "gpt2-xl": (["input_ids"], 11, True, "gpt2"),
-    "distilgpt2": (["input_ids"], 11, False, "gpt2"),
-    # Transformer-XL (Models uses Einsum, which need opset version 12 or later.)
-    "transfo-xl-wt103": (["input_ids", "mems"], 12, False, "bert"),
+    "bert-base-cased": (["input_ids", "attention_mask", "token_type_ids"], 16, False, "bert"),
+    "bert-large-cased": (["input_ids", "attention_mask", "token_type_ids"], 16, False, "bert"),
+    # Transformer-XL (Models uses Einsum, which need opset version 16 or later.)
+    "transfo-xl-wt103": (["input_ids", "mems"], 16, False, "bert"),
     # XLNet
-    "xlnet-base-cased": (["input_ids"], 12, False, "bert"),
-    "xlnet-large-cased": (["input_ids"], 12, False, "bert"),
+    "xlnet-base-cased": (["input_ids"], 16, False, "bert"),
+    "xlnet-large-cased": (["input_ids"], 16, False, "bert"),
     # XLM
-    "xlm-mlm-en-2048": (["input_ids"], 11, True, "bert"),
-    "xlm-mlm-ende-1024": (["input_ids"], 11, False, "bert"),
-    "xlm-mlm-enfr-1024": (["input_ids"], 11, False, "bert"),
+    "xlm-mlm-en-2048": (["input_ids"], 16, True, "bert"),
+    "xlm-mlm-ende-1024": (["input_ids"], 16, False, "bert"),
+    "xlm-mlm-enfr-1024": (["input_ids"], 16, False, "bert"),
     # RoBERTa
-    "roberta-base": (["input_ids", "attention_mask"], 12, False, "bert"),
-    "roberta-large": (["input_ids", "attention_mask"], 12, False, "bert"),
-    "roberta-large-mnli": (["input_ids", "attention_mask"], 12, False, "bert"),
-    "deepset/roberta-base-squad2": (["input_ids", "attention_mask"], 11, False, "bert"),
-    "distilroberta-base": (["input_ids", "attention_mask"], 12, False, "bert"),
+    "roberta-base": (["input_ids", "attention_mask"], 16, False, "bert"),
+    "roberta-large": (["input_ids", "attention_mask"], 16, False, "bert"),
+    "roberta-large-mnli": (["input_ids", "attention_mask"], 16, False, "bert"),
+    "deepset/roberta-base-squad2": (["input_ids", "attention_mask"], 16, False, "bert"),
+    "distilroberta-base": (["input_ids", "attention_mask"], 16, False, "bert"),
     # DistilBERT
-    "distilbert-base-uncased": (["input_ids", "attention_mask"], 11, False, "bert"),
-    "distilbert-base-uncased-distilled-squad": (
-        ["input_ids", "attention_mask"],
-        11,
-        False,
-        "bert",
-    ),
+    "distilbert-base-uncased": (["input_ids", "attention_mask"], 16, False, "bert"),
+    "distilbert-base-uncased-distilled-squad": (["input_ids", "attention_mask"], 16, False, "bert"),
     # CTRL
-    "ctrl": (["input_ids"], 11, True, "bert"),
+    "ctrl": (["input_ids"], 16, True, "bert"),
     # CamemBERT
-    "camembert-base": (["input_ids"], 11, False, "bert"),
+    "camembert-base": (["input_ids"], 16, False, "bert"),
     # ALBERT
-    "albert-base-v1": (["input_ids"], 12, False, "bert"),
-    "albert-large-v1": (["input_ids"], 12, False, "bert"),
-    "albert-xlarge-v1": (["input_ids"], 12, True, "bert"),
-    # "albert-xxlarge-v1": (["input_ids"], 12, True, "bert"),
-    "albert-base-v2": (["input_ids"], 12, False, "bert"),
-    "albert-large-v2": (["input_ids"], 12, False, "bert"),
-    "albert-xlarge-v2": (["input_ids"], 12, True, "bert"),
-    # "albert-xxlarge-v2": (["input_ids"], 12, True, "bert"),
-    # T5 (use benchmark_t5.py instead)
-    # "t5-small": (["input_ids", "decoder_input_ids"], 12, False, "bert"),
-    # "t5-base": (["input_ids", "decoder_input_ids"], 12, False, "bert"),
-    # "t5-large": (["input_ids", "decoder_input_ids"], 12, True, "bert"),
-    # "t5-3b": (["input_ids", "decoder_input_ids"], 12, True, "bert"),
-    # "t5-11b": (["input_ids", "decoder_input_ids"], 12, True, "bert"),
-    # "valhalla/t5-small-qa-qg-hl": (["input_ids"], 12, True, "bert"),
+    "albert-base-v1": (["input_ids"], 16, False, "bert"),
+    "albert-large-v1": (["input_ids"], 16, False, "bert"),
+    "albert-xlarge-v1": (["input_ids"], 16, True, "bert"),
+    # "albert-xxlarge-v1": (["input_ids"], 16, True, "bert"),
+    "albert-base-v2": (["input_ids"], 16, False, "bert"),
+    "albert-large-v2": (["input_ids"], 16, False, "bert"),
+    "albert-xlarge-v2": (["input_ids"], 16, True, "bert"),
+    # "albert-xxlarge-v2": (["input_ids"], 16, True, "bert"),
     # XLM-RoBERTa
-    "xlm-roberta-base": (["input_ids"], 11, False, "bert"),
-    "xlm-roberta-large": (["input_ids"], 11, True, "bert"),
+    "xlm-roberta-base": (["input_ids"], 16, False, "bert"),
+    "xlm-roberta-large": (["input_ids"], 16, True, "bert"),
     # FlauBERT
-    "flaubert/flaubert_small_cased": (["input_ids"], 11, False, "bert"),
-    # "flaubert/flaubert_base_uncased": (["input_ids"], 11, False, "bert"),
-    "flaubert/flaubert_base_cased": (["input_ids"], 11, False, "bert"),
-    # "flaubert/flaubert_large_cased": (["input_ids"], 11, False, "bert"),
-    # Bart
-    "facebook/bart-large": (["input_ids", "attention_mask"], 11, False, "bart"),
-    "facebook/bart-base": (["input_ids", "attention_mask"], 11, False, "bart"),
-    "facebook/bart-large-mnli": (["input_ids", "attention_mask"], 11, False, "bart"),
-    "facebook/bart-large-cnn": (["input_ids", "attention_mask"], 11, False, "bart"),
-    # DialoGPT
-    "microsoft/DialoGPT-small": (["input_ids"], 11, False, "gpt2"),
-    "microsoft/DialoGPT-medium": (["input_ids"], 11, False, "gpt2"),
-    # "microsoft/DialoGPT-large": (["input_ids"], 11, True, "gpt2"),
-    # Reformer
-    # "google/reformer-enwik8": (["input_ids"], 11, False, "bert"),
-    # "google/reformer-crime-and-punishment": (["input_ids"], 11, False, "bert"),
-    # MarianMT
-    # "Helsinki-NLP/opus-mt-ROMANCE-en": (["input_ids"], 12, False, "bert"),
-    # Longformer (use benchmark_longformer.py instead)
-    # "allenai/longformer-base-4096": (["input_ids"], 12, False, "bert"),
-    # "allenai/longformer-large-4096": (["input_ids"], 12, False, "bert"),
-    # MBart
-    "facebook/mbart-large-cc25": (["input_ids"], 11, True, "bert"),
-    "facebook/mbart-large-en-ro": (["input_ids"], 11, True, "bert"),
-    # "Helsinki-NLP/opus-mt-ROMANCE-en": (["input_ids"], 12, False, "bert"),
-    # # Longformer
-    # "allenai/longformer-base-4096": (["input_ids"], 12, False, "bert"),
-    # "allenai/longformer-large-4096": (["input_ids"], 12, True, "bert"),
-    # "funnel-transformer/small": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/small-base": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/medium": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/medium-base": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/intermediate": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/intermediate-base": (["input_ids"], 12, False, "bert"),
-    # "funnel-transformer/large": (["input_ids"], 12, True, "bert"),
-    # "funnel-transformer/large-base": (["input_ids"], 12, True, "bert"),
-    # "funnel-transformer/xlarge": (["input_ids"], 12, True, "bert"),
-    # "funnel-transformer/xlarge-base": (["input_ids"], 12, True, "bert"),
+    "flaubert/flaubert_small_cased": (["input_ids"], 16, False, "bert"),
+    "flaubert/flaubert_base_cased": (["input_ids"], 16, False, "bert"),
+    # "flaubert/flaubert_large_cased": (["input_ids"], 16, False, "bert"),
     # Layoutlm
-    "microsoft/layoutlm-base-uncased": (["input_ids"], 11, False, "bert"),
-    "microsoft/layoutlm-large-uncased": (["input_ids"], 11, False, "bert"),
+    "microsoft/layoutlm-base-uncased": (["input_ids"], 16, False, "bert"),
+    "microsoft/layoutlm-large-uncased": (["input_ids"], 16, False, "bert"),
     # Squeezebert
-    "squeezebert/squeezebert-uncased": (["input_ids"], 11, False, "bert"),
-    "squeezebert/squeezebert-mnli": (["input_ids"], 11, False, "bert"),
-    "squeezebert/squeezebert-mnli-headless": (["input_ids"], 11, False, "bert"),
-    "unc-nlp/lxmert-base-uncased": (
-        ["input_ids", "visual_feats", "visual_pos"],
-        11,
-        False,
-        "bert",
-    ),
-    # "google/pegasus-xsum": (["input_ids"], 11, False, "bert"),
-    # "google/pegasus-large": (["input_ids"], 11, False, "bert"),
+    "squeezebert/squeezebert-uncased": (["input_ids"], 16, False, "bert"),
+    "squeezebert/squeezebert-mnli": (["input_ids"], 16, False, "bert"),
+    "squeezebert/squeezebert-mnli-headless": (["input_ids"], 16, False, "bert"),
+    "unc-nlp/lxmert-base-uncased": (["input_ids", "visual_feats", "visual_pos"], 16, False, "bert"),
     # ViT
-    "google/vit-base-patch16-224": (["pixel_values"], 12, False, "vit"),
+    "google/vit-base-patch16-224": (["pixel_values"], 16, False, "vit"),
     # Swin
-    "microsoft/swin-base-patch4-window7-224": (["pixel_values"], 12, False, "swin"),
-    "microsoft/swin-small-patch4-window7-224": (["pixel_values"], 12, False, "swin"),
-    "microsoft/swin-tiny-patch4-window7-224": (["pixel_values"], 12, False, "swin"),
+    "microsoft/swin-base-patch4-window7-224": (["pixel_values"], 16, False, "swin"),
+    "microsoft/swin-small-patch4-window7-224": (["pixel_values"], 16, False, "swin"),
+    "microsoft/swin-tiny-patch4-window7-224": (["pixel_values"], 16, False, "swin"),
 }
diff --git a/onnxruntime/python/tools/transformers/onnx_exporter.py b/onnxruntime/python/tools/transformers/onnx_exporter.py
@@ -492,7 +492,7 @@ def export_onnx_model_from_pt(
         example_inputs = image_processor(data, return_tensors="pt")
     else:
         tokenizer = AutoTokenizer.from_pretrained(model_name, cache_dir=cache_dir)
-        max_input_size = tokenizer.max_model_input_sizes.get(model_name, 1024)
+        max_input_size = tokenizer.model_max_length
         example_inputs = tokenizer.encode_plus("This is a sample input", return_tensors="pt")
 
     example_inputs = filter_inputs(example_inputs, input_names)
@@ -596,7 +596,7 @@ def export_onnx_model_from_tf(
     # Fix "Using pad_token, but it is not set yet" error.
     if tokenizer.pad_token is None:
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
-    max_input_size = tokenizer.max_model_input_sizes.get(model_name, 1024)
+    max_input_size = tokenizer.model_max_length
 
     config, model = load_tf_model(model_name, model_class, cache_dir, config_modifier)
     model.resize_token_embeddings(len(tokenizer))
diff --git a/onnxruntime/python/tools/transformers/run_benchmark.sh b/onnxruntime/python/tools/transformers/run_benchmark.sh
@@ -5,10 +5,7 @@
 # license information.
 # --------------------------------------------------------------------------
 # This measures the performance of OnnxRuntime, PyTorch and TorchScript on transformer models.
-# Please install PyTorch (see https://pytorch.org/) before running this benchmark. Like the following:
-# GPU:   conda install pytorch torchvision cudatoolkit=11.0 -c pytorch
-# CPU:   conda install pytorch torchvision cpuonly -c pytorch
-# To use torch2, please install the nightly PyTorch by replacing pytorch with pytorch-nightly.
+# Please install PyTorch (see https://pytorch.org/) before running this benchmark.
 
 # When use_package=true, you need not copy other files to run benchmarks except this sh file.
 # Otherwise, it will use python script (*.py) files in this directory.
@@ -60,7 +57,6 @@ sequence_lengths="8 16 32 64 128 256 512 1024"
 # Here we only test one input (input_ids) for fair comparison with PyTorch.
 input_counts=1
 
-# Pretrained transformers models can be a subset of: bert-base-cased roberta-base gpt2 distilgpt2 distilbert-base-uncased
 models_to_test="bert-base-cased roberta-base distilbert-base-uncased"
 
 # If you have multiple GPUs, you can choose one GPU for test. Here is an example to use the second GPU:
@@ -99,7 +95,7 @@ if [ "$run_install" = true ] ; then
   else
     pip install onnxruntime-gpu
   fi
-  pip install --upgrade onnx coloredlogs packaging psutil py3nvml onnxconverter_common numpy transformers sympy
+  pip install --upgrade onnx coloredlogs packaging psutil py3nvml numpy transformers sympy
 fi
 
 if [ "$use_package" = true ] ; then
