Method _encode_pair

lm-eval-harness/lm_eval/models/huggingface.py:814–862

(
        self, context: str, continuation: str, decode_mode="default"
    )

        return logits

    def _encode_pair(
        self, context: str, continuation: str, decode_mode="default"
    ) -> Tuple[List[int], List[int]]:
        
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]

        whole_enc = self.tok_encode(context + continuation, add_special_tokens=False)
        context_enc = self.tok_encode(context, add_special_tokens=False)

        def custom_enc_dec_encode(
            encodings,
            whole_enc, 
            pad_token_id: int = 220, #50256,
            enc_length: int=1024,
        ): 
            encodings = torch.tensor(encodings).unsqueeze(0)
            whole_enc = torch.tensor(whole_enc).unsqueeze(0)
            batch, seqlen = whole_enc.shape
            if seqlen < enc_length:
                prefill_size_pad = enc_length - seqlen 
                padding = torch.full((batch, prefill_size_pad), pad_token_id, dtype=torch.long, device=encodings.device)
                encodings = torch.cat([padding, encodings], dim=-1)
                whole_enc = torch.cat([padding, whole_enc], dim=-1)
                attn_mask = torch.ones_like(whole_enc)
                attn_mask[:, :prefill_size_pad] = 0

                # convert back to list
                encodings = encodings.squeeze(0).tolist()
                whole_enc = whole_enc.squeeze(0).tolist()
                return encodings, whole_enc, attn_mask, enc_length
            else:
                # attn_mask = torch.ones_like(whole_enc)
                attn_mask=None
                encodings = encodings.squeeze(0).tolist()
                whole_enc = whole_enc.squeeze(0).tolist()
                return encodings, whole_enc, attn_mask, seqlen
        if decode_mode == "default_left_pad":
            context_enc, whole_enc, attn_mask, seqlen = custom_enc_dec_encode(context_enc, whole_enc)
        else:
            attn_mask = None

        # whole_enc = self.tok_encode(context + continuation)
        # context_enc = self.tok_encode(context, add_special_tokens=False)
        context_enc_len = len(context_enc)
        continuation_enc = whole_enc[context_enc_len:]
        return context_enc, continuation_enc, attn_mask
    
    def loglikelihood(self, requests: List[Instance], decode_mode="default") -> List[Tuple[float, bool]]:
        new_reqs = []