llama v2 (#1775)

Summary: Pull Request resolved: #1775

Reviewed By: xuzhao9

Differential Revision: D47723915

Pulled By: msaroufim

fbshipit-source-id: ef6686b328f691833907660ded2fbb28d317ea61

requirements.txt

@@ -10,7 +10,7 @@ pytest-benchmark
 requests
 tabulate
 git+https://github.com/huggingface/pytorch-image-models.git@6635bc3
-transformers==4.30.0
+transformers==4.31.0
 MonkeyType
 psutil
 pyyaml

torchbenchmark/models/llama_v2_7b_16h/__init__.py

@@ -0,0 +1,15 @@
+from torchbenchmark.tasks import NLP
+from torchbenchmark.util.framework.huggingface.model_factory import HuggingFaceModel, HuggingFaceAuthMixin
+
+class Model(HuggingFaceModel, HuggingFaceAuthMixin):
+    task = NLP.LANGUAGE_MODELING
+    DEFAULT_TRAIN_BSIZE = 1
+    DEFAULT_EVAL_BSIZE = 1
+    DEEPCOPY = False 
+
+    def __init__(self, test, device, jit=False, batch_size=None, extra_args=[]):
+        HuggingFaceAuthMixin.__init__(self)
+        super().__init__(name="llama_v2", test=test, device=device, jit=jit, batch_size=batch_size, extra_args=extra_args)
+
+    def train(self):
+        return NotImplementedError("7b LLAMA model will OOM on CI GPU machines")

torchbenchmark/models/llama_v2_7b_16h/install.py

@@ -0,0 +1,9 @@
+
+import subprocess
+import sys
+import os
+from torchbenchmark.util.framework.huggingface.patch_hf import patch_transformers, cache_model
+
+if __name__ == '__main__':
+    model_name = os.path.basename(os.path.dirname(os.path.abspath(__file__)))
+    cache_model(model_name)

torchbenchmark/models/llama_v2_7b_16h/metadata.yaml

@@ -0,0 +1,13 @@
+devices:
+  NVIDIA A100-SXM4-40GB:
+    eval_batch_size: 1
+eval_benchmark: false
+eval_deterministic: false
+eval_nograd: true
+not_implemented:
+- jit: true
+- device: cpu
+- device: cuda
+  test: train
+train_benchmark: false
+train_deterministic: false

torchbenchmark/util/framework/huggingface/model_factory.py

@@ -8,7 +8,7 @@
 from torchbenchmark.util.model import BenchmarkModel
 from torchbenchmark.tasks import NLP
 import transformers
-from transformers import AutoConfig, ReformerConfig, BertConfig, GenerationConfig
+from transformers import AutoConfig, ReformerConfig, BertConfig, LlamaConfig, GenerationConfig
 from typing import Tuple
 
 class_models = {
@@ -27,6 +27,8 @@
     'hf_Bert': (512, 512, 'BertConfig()', 'AutoModelForMaskedLM'),
     # see https://huggingface.co/bert-large-cased
     'hf_Bert_large': (512, 512, 'BertConfig(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16)', 'AutoModelForMaskedLM'),
+    # default num_hidden_layers=32 but that OOMs, feel free to change this config to something more real
+    'llama_v2_7b_16h' : (512,512, 'LlamaConfig(num_hidden_layers=16)', 'AutoModelForCausalLM'),
 }
 
 cpu_input_slice = {

torchbenchmark/util/framework/huggingface/patch_hf.py

@@ -5,7 +5,7 @@
 import subprocess
 import sys
 from .model_factory import class_models
-from transformers import AutoConfig, ReformerConfig, BigBirdConfig, BertConfig
+from transformers import AutoConfig, ReformerConfig, BigBirdConfig, BertConfig, LlamaConfig
 
 PATCH_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "patches")