gpt2.py

import transformers
import torch
import torch.nn.functional as F
from lm_eval.base import LM
from lm_eval import utils
from tqdm import tqdm


class GPT2LM(LM):
    def __init__(self, device="cpu"):
        self.device = torch.device(device)
        self.gpt2 = transformers.GPT2LMHeadModel.from_pretrained('gpt2').to(self.device)
        self.gpt2.eval()
        self.tokenizer = transformers.GPT2TokenizerFast.from_pretrained('gpt2')

    @classmethod
    def create_from_arg_string(cls, arg_string):
        args = utils.simple_parse_args_string(arg_string)
        return cls(device=args.get("device", "cpu"))

    def loglikelihood(self, requests):
        res = []
        # TODO: vectorize properly
        for context, continuation in tqdm(requests):
            # when too long to fit in context, truncate from the left
            context_enc = self.tokenizer.encode(context)
            continuation_enc = self.tokenizer.encode(continuation)
            inp = torch.tensor([(context_enc + continuation_enc)[-1024:]], dtype=torch.long).to(self.device)
            ctxlen = len(context_enc) - max(0, len(context_enc) + len(continuation_enc) - 1024)

            cont_toks = inp[:, ctxlen:]  # [batch, seq]
            logits = F.log_softmax(self.gpt2(inp)[0], dim=-1)[:, ctxlen - 1:-1]  # [batch, seq, vocab]
            
            greedy_tokens = logits.argmax(dim=-1)
            max_equal = (greedy_tokens == cont_toks).all()

            logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(-1) # [batch, seq]


            res.append((float(logits.sum()), bool(max_equal)))

        return res
    
    def greedy_until(self, requests):
        # TODO: implement
        pass