Benchmarks - Add LLaMA-2 Models

docs/superbench-config.mdx

@@ -328,7 +328,8 @@ A list of models to run, only supported in model-benchmark.
     shufflenet_v2_x0_5 | shufflenet_v2_x1_0 | shufflenet_v2_x1_5 | shufflenet_v2_x2_0 |
     squeezenet1_0 | squeezenet1_1 |
     vgg11 | vgg11_bn | vgg13 | vgg13_bn | vgg16 | vgg16_bn | vgg19_bn | vgg19 |
-    bert-base | bert-large | gpt2-small | gpt2-medium | gpt2-large | gpt2-xl ]
+    bert-base | bert-large | gpt2-small | gpt2-medium | gpt2-large | gpt2-xl |
+    llama2-7b | llama2-13b | llama2-70b ]
   ```
 * default value: `[ ]`
 

docs/user-tutorial/benchmarks/model-benchmarks.md

@@ -13,6 +13,7 @@ id: model-benchmarks
 Run training or inference tasks with single or half precision for deep learning models,
 including the following categories:
 * GPT: gpt2-small, gpt2-medium, gpt2-large and gpt2-xl
+* LLAMA: llama2-7b, llama2-13b, llama2-70b
 * BERT: bert-base and bert-large
 * LSTM
 * CNN, listed in [`torchvision.models`](https://pytorch.org/vision/0.8/models.html), including:

examples/benchmarks/pytorch_llama2.py

@@ -0,0 +1,41 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+"""Model benchmark example for Llama2-7b (32-layer, 4096-hidden, 32-heads, 7B parameters).
+
+Commands to run:
+  python3 examples/benchmarks/pytorch_llama2.py (Single GPU)
+  python3 -m torch.distributed.launch --use_env --nproc_per_node=8 examples/benchmarks/pytorch_llama2.py \
+      --distributed (Distributed)
+"""
+
+import argparse
+
+from superbench.benchmarks import Platform, Framework, BenchmarkRegistry
+from superbench.common.utils import logger
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--distributed', action='store_true', default=False, help='Whether to enable distributed training.'
+    )
+    args = parser.parse_args()
+
+    # Specify the model name and benchmark parameters.
+    model_name = 'llama2-7b'
+    parameters = '--batch_size 1 --duration 120 --seq_len 512 --precision float16'
+    if args.distributed:
+        parameters += ' --distributed_impl ddp --distributed_backend nccl'
+
+    # Create context for Llama2 benchmark and run it for 120 seconds.
+    context = BenchmarkRegistry.create_benchmark_context(
+        model_name, platform=Platform.CUDA, parameters=parameters, framework=Framework.PYTORCH
+    )
+
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+    if benchmark:
+        logger.info(
+            'benchmark: {}, return code: {}, result: {}'.format(
+                benchmark.name, benchmark.return_code, benchmark.result
+            )
+        )

setup.py

@@ -209,9 +209,10 @@ def run(self):
                 'yapf==0.31.0',
             ],
             'torch': [
+                'tokenizers<=0.20.3',
                 'torch>=1.7.0a0',
                 'torchvision>=0.8.0a0',
-                'transformers>=4.3.3, <4.23.0',
+                'transformers>=4.28.0',
             ],
             'ort': [
                 'onnx>=1.10.2',

superbench/benchmarks/base.py

@@ -89,7 +89,8 @@ def get_configurable_settings(self):
         Return:
             All configurable settings in raw string.
         """
-        return self._parser.format_help().strip()
+        message = self._parser.format_help().strip()
+        return message
 
     def parse_args(self, ignore_invalid=False):
         """Parse the arguments.

superbench/benchmarks/micro_benchmarks/_export_torch_to_onnx.py

@@ -9,11 +9,12 @@
 import torch.hub
 import torch.onnx
 import torchvision.models
-from transformers import BertConfig, GPT2Config
+from transformers import BertConfig, GPT2Config, LlamaConfig
 
 from superbench.benchmarks.model_benchmarks.pytorch_bert import BertBenchmarkModel
 from superbench.benchmarks.model_benchmarks.pytorch_gpt2 import GPT2BenchmarkModel
 from superbench.benchmarks.model_benchmarks.pytorch_lstm import LSTMBenchmarkModel
+from superbench.benchmarks.model_benchmarks.pytorch_llama import LlamaBenchmarkModel
 
 
 class torch2onnxExporter():
@@ -87,6 +88,39 @@ def __init__(self):
                 ),
                 self.num_classes,
             ),
+            'llama2-7b':
+            lambda: LlamaBenchmarkModel(
+                LlamaConfig(
+                    hidden_size=4096,
+                    num_hidden_layers=32,
+                    num_attention_heads=32,
+                    num_key_value_heads=32,
+                    intermediate_size=11008,
+                ),
+                self.num_classes,
+            ),
+            'llama2-13b':
+            lambda: LlamaBenchmarkModel(
+                LlamaConfig(
+                    hidden_size=5120,
+                    num_hidden_layers=40,
+                    num_attention_heads=40,
+                    num_key_value_heads=40,
+                    intermediate_size=13824,
+                ),
+                self.num_classes,
+            ),
+            'llama2-70b':
+            lambda: LlamaBenchmarkModel(
+                LlamaConfig(
+                    hidden_size=8192,
+                    num_hidden_layers=80,
+                    num_attention_heads=64,
+                    num_key_value_heads=8,
+                    intermediate_size=28672,
+                ),
+                self.num_classes,
+            ),
         }
         self._onnx_model_path = Path(torch.hub.get_dir()) / 'onnx'
         self._onnx_model_path.mkdir(parents=True, exist_ok=True)
@@ -138,7 +172,7 @@ def export_torchvision_model(self, model_name, batch_size=1):
             model,
             dummy_input,
             file_name,
-            opset_version=10,
+            opset_version=14,
             operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK,
             input_names=['input'],
             output_names=['output'],
@@ -179,7 +213,7 @@ def export_benchmark_model(self, model_name, batch_size=1, seq_length=512):
             model,
             dummy_input,
             file_name,
-            opset_version=10,
+            opset_version=14,
             do_constant_folding=True,
             input_names=['input'],
             output_names=['output'],

superbench/benchmarks/model_benchmarks/__init__.py

@@ -10,4 +10,4 @@
 from superbench.benchmarks.model_benchmarks.pytorch_lstm import PytorchLSTM
 from superbench.benchmarks.model_benchmarks.megatron_gpt3 import MegatronGPT
 
-__all__ = ['ModelBenchmark', 'PytorchBERT', 'PytorchGPT2', 'PytorchCNN', 'PytorchLSTM', 'MegatronGPT']
+__all__ = ['ModelBenchmark', 'PytorchBERT', 'PytorchGPT2', 'PytorchCNN', 'PytorchLSTM', 'MegatronGPT', 'PytorchLlama']