gpt2.py

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


class GPT2LM(LM):
    MAX_GEN_TOKS = 256
    VOCAB_SIZE = 50257
    EOT_TOKEN_ID = 50256

    def __init__(self, device='cuda', pretrained='gpt2', batch_size=1):
        super().__init__()
        if device:
            self.device = torch.device(device)
        else:
            self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
        self.gpt2 = transformers.AutoModelForCausalLM.from_pretrained(pretrained).to(self.device)
        self.gpt2.eval()

        # pretrained tokenizer for neo is broken for now so just hardcoding this to gpt2
        self.tokenizer = transformers.GPT2TokenizerFast.from_pretrained('gpt2')
        self.tokenizer.pad_token = "<|endoftext|>"
        try:
            self.max_length = self.gpt2.config.n_ctx
        except AttributeError:
            # gptneoconfig doesn't have n_ctx apparantly
            self.max_length = self.gpt2.config.max_position_embeddings

        assert self.tokenizer.encode('hello\n\nhello') == [31373, 198, 198, 31373]

        # multithreading and batching
        gpus = torch.cuda.device_count()
        batch_size_per_gpu = batch_size # todo: adaptive batch size

        self.batch_size = batch_size_per_gpu * gpus

        # TODO: fix multi-gpu
        # if gpus > 1:
        #     self.gpt2 = nn.DataParallel(self.gpt2)

    @classmethod
    def create_from_arg_string(cls, arg_string, **kwargs):
        args = utils.simple_parse_args_string(arg_string)
        kwargs = {k: v for k, v in kwargs.items() if v is not None}
        return cls(pretrained=args.get("pretrained", "gpt2"), **kwargs)

    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in requests:
            if context == "":
                # end of text as context
                context_enc = [self.EOT_TOKEN_ID]
            else:
                context_enc = self.tokenizer.encode(context)

            continuation_enc = self.tokenizer.encode(continuation)

            new_reqs.append(((context, continuation), context_enc, continuation_enc))

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        # TODO: Implement caching once we've confirmed the perplexity implementation
        # TODO: automatic batch size detection for vectorization

        loglikelihoods = []
        with torch.no_grad():
            for string, in tqdm(requests):
                rolling_token_windows = list(map(utils.make_disjoint_window, utils.get_rolling_token_windows(
                    token_list=self.tokenizer.encode(string),
                    prefix_token=self.EOT_TOKEN_ID,
                    max_seq_len=self.max_length,
                    context_len=1,
                )))

                rolling_token_windows = [(None,) + x for x in rolling_token_windows]

                # TODO: extract out this call so it only gets called once and also somehow figure out partial caching for that
                string_nll = self._loglikelihood_tokens(rolling_token_windows, disable_tqdm=True)
                
                # discard is_greedy
                string_nll = [x[0] for x in string_nll]
                
                string_nll = sum(string_nll)
                loglikelihoods.append(string_nll)

        return loglikelihoods

    def _loglikelihood_tokens(self, requests, disable_tqdm=False):
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []
        with torch.no_grad():

            def _collate(x):
                # the negative sign on len(toks) sorts descending - this has a few advantages:
                # - time estimates will always be over not underestimates, which is more useful for planning
                # - to know the size of a batch when going through the list, you know the first one is always the batch padded context length.
                #   this is useful to simplify the batching logic and more importantly to make automatic adaptive batches much much easier to implement
                # - any OOMs will happen right away rather than near the end

                toks = x[1] + x[2]
                return (-len(toks), tuple(toks))
            
            # TODO: automatic (variable) batch size detection for vectorization
            reord = utils.Reorderer(requests, _collate)
            for chunk in utils.chunks(tqdm(reord.get_reordered(), disable=disable_tqdm), self.batch_size):
                inps = []
                contlens = []
                inplens = []

                padding_length = None

                # because vectorizing is annoying, we first convert each (context, continuation) pair to padded
                # tensors, then we pack them together into a batch, call the model, and then pick it all apart
                # again because vectorizing is annoying

                for _, context_enc, continuation_enc in chunk:
                    # sanity check
                    assert len(context_enc) > 0
                    assert len(continuation_enc) > 0
                    assert len(continuation_enc) <= self.max_length

                    # how this all works:
                    #          CTX      CONT
                    # inp    0 1 2 3|4 5 6 7 8 9 <- last token is deleted by inp[:, :-1]
                    # gpt2    \               \
                    # logits   1 2 3|4 5 6 7 8 9   <- the ctx half gets tossed out by the [:, -len(continuation_enc):, :self.VOCAB_SIZE] slice
                    # cont_toks      4 5 6 7 8 9

                    # when too long to fit in context, truncate from the left
                    inp = torch.tensor(
                        (context_enc + continuation_enc)[-(self.max_length+1):][:-1]
                    , dtype=torch.long).to(self.device)
                    inplen, = inp.shape

                    cont = continuation_enc

                    # since in _collate we make sure length is descending, the longest is always the first one.
                    padding_length = padding_length if padding_length is not None else inplen

                    # pad to length
                    inp = torch.cat([
                        inp, # [seq]
                        torch.zeros(padding_length - inplen, dtype=torch.long).to(inp.device) # [padding_length - seq]
                    ], dim=0)

                    inps.append(inp.unsqueeze(0))
                    contlens.append(cont)
                    inplens.append(inplen)

                multi_logits = F.log_softmax(self._model_call(torch.cat(inps, dim=0)), dim=-1).cpu()  # [batch, seq, vocab]

                for (cache_key, _, _), logits, inp, inplen, cont_toks in zip(chunk, multi_logits, inps, inplens, contlens):
                    contlen = len(cont_toks)

                    logits = logits[inplen-contlen:inplen].unsqueeze(0) # [1, seq, vocab]

                    greedy_tokens = logits.argmax(dim=-1)

                    # cont_toks :: [1, seq]
                    cont_toks = torch.tensor(cont_toks, dtype=torch.long).unsqueeze(0)

                    max_equal = (greedy_tokens == cont_toks).all()

                    #last_token_slice = logits[:, -1, :].squeeze(0).tolist()

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

                    answer = (float(logits.sum()), bool(max_equal))

                    # partial caching
                    if cache_key is not None:
                        self.cache_hook.add_partial("loglikelihood", cache_key, answer)

                    res.append(answer)

        return reord.get_original(res)
    
    def _model_call(self, inps):
        """
        inps: a torch tensor of shape [batch, sequence]
        the size of sequence may vary from call to call

        returns: a torch tensor of shape [batch, sequence, vocab] with the
        logits retuned from the model
        """
        return self.gpt2(inps)[0][:, :, :50257]
    
    def greedy_until(self, requests):
        # TODO: implement fully general `until` that handles untils that are 
        # multiple tokens or that span multiple tokens correctly
        res = []

        def _collate(x):
            toks = self.tokenizer.encode(x[0])
            return (len(toks), x[0])
        
        reord = utils.Reorderer(requests, _collate)

        for context, until in tqdm(reord.get_reordered()):
            if isinstance(until, str): until = [until]

            context_enc = torch.tensor([self.tokenizer.encode(context)[self.MAX_GEN_TOKS - self.max_length:]]).to(self.device)

            primary_until, = self.tokenizer.encode(until[0])

            cont = self.gpt2.generate(
                context_enc,
                max_length=context_enc.shape[1] + self.MAX_GEN_TOKS,
                eos_token_id=primary_until,
                do_sample=False
            )

            s = self.tokenizer.decode(cont[0].tolist()[context_enc.shape[1]:])

            for term in until:
                s = s.split(term)[0]
            
            # partial caching
            self.cache_hook.add_partial("greedy_until", (context, until), s)
            
            res.append(s)
        
        return reord.get_original(res)