Disable llama_v2_7b_16h on optim benchmarks as they currently OOM

torchbenchmark/models/llama_v2_7b_8h/__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, batch_size=None, extra_args=[]):
+        HuggingFaceAuthMixin.__init__(self)
+        super().__init__(name="llama_v2_7b_8h", test=test, device=device, 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_8h/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_8h/metadata.yaml

@@ -0,0 +1,12 @@
+devices:
+  NVIDIA A100-SXM4-40GB:
+    eval_batch_size: 1
+eval_benchmark: false
+eval_deterministic: false
+eval_nograd: true
+not_implemented:
+- device: cpu
+- device: cuda
+  test: train
+train_benchmark: false
+train_deterministic: false

torchbenchmark/util/framework/huggingface/model_factory.py

@@ -30,6 +30,8 @@
     'hf_Whisper': (1024, 1024, 'WhisperConfig()', 'AutoModelForAudioClassification'),
     # 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'),
+    # an even smaller model since 16h OOMs for our optimizer benchmarks
+    'llama_v2_7b_8h' : (512,512, 'LlamaConfig(num_hidden_layers=8)', 'AutoModelForCausalLM'),
 }
 
 cpu_input_slice = {