Added codellama (testing, in canary).

torchbenchmark/canary_models/codellama/__init__.py

@@ -0,0 +1,16 @@
+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="codellama", test=test, device=device, batch_size=batch_size, extra_args=extra_args)
+
+
+    # def train(self):
+    #     return NotImplementedError("FSDP should implement a training loop")

torchbenchmark/canary_models/codellama/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/canary_models/codellama/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

@@ -36,6 +36,7 @@
     'llama_v2_7b' : (512,512, 'AutoConfig.from_pretrained("meta-llama/Llama-2-7b-hf")', 'AutoModelForCausalLM'),
     'llama_v2_13b' : (512,512, 'AutoConfig.from_pretrained("meta-llama/Llama-2-13b-hf")', 'AutoModelForCausalLM'),
     'llama_v2_70b' : (512, 512, 'AutoConfig.from_pretrained("meta-llama/Llama-2-70b-hf")', 'AutoModelForMaskedLM'),
+    'codellama' : (512,512, 'AutoConfig.from_pretrained("codellama/CodeLlama-7b-hf")', 'AutoModelForCausalLM'),
     'phi_1_5' : (512, 512, 'AutoConfig.from_pretrained("microsoft/phi-1_5", trust_remote_code=True)', 'AutoModelForCausalLM'),
     'phi_2' : (512, 512, 'AutoConfig.from_pretrained("microsoft/phi-2", trust_remote_code=True)', 'AutoModelForCausalLM'),
     # as per this page https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1 trust_remote_code=True is not required