Update gpt2 for efficiency and allow specifying model size

lm_eval/models/gpt2.py

@@ -7,20 +7,21 @@ from tqdm import tqdm
 
 
 class GPT2LM(LM):
-    def __init__(self, device="cpu"):
+    def __init__(self, device="cpu", pretrained='gpt2'):
         self.device = torch.device(device)
-        self.gpt2 = transformers.GPT2LMHeadModel.from_pretrained('gpt2').to(self.device)
+        self.gpt2 = transformers.GPT2LMHeadModel.from_pretrained(pretrained).to(self.device)
         self.gpt2.eval()
-        self.tokenizer = transformers.GPT2TokenizerFast.from_pretrained('gpt2')
+        self.tokenizer = transformers.GPT2TokenizerFast.from_pretrained(pretrained)
         self.tokenizer.pad_token = "<|endoftext|>"
 
     @classmethod
     def create_from_arg_string(cls, arg_string):
         args = utils.simple_parse_args_string(arg_string)
-        return cls(device=args.get("device", "cpu"))
+        return cls(device=args.get("device", "cpu"), pretrained=args.get("pretrained", "gpt2"))
 
     def loglikelihood(self, requests):
         res = []
+        with torch.no_grad():
             # TODO: vectorize properly
             for context, continuation in tqdm(requests):
                 # when too long to fit in context, truncate from the left