Merge pull request #5 from EleutherAI/seq2seq-support

fix edge cases for seq2seq

lm_eval/models/huggingface.py

@@ -127,7 +127,7 @@ class HFLM(LM):
                 )
                 self.model.to(self.device)
             else:
-                self.model = accelerator.prepare(self.model)
+                self._model = accelerator.prepare(self.model)
                 self._device = torch.device(f"cuda:{accelerator.local_process_index}")
                 self.accelerator = accelerator
 
@@ -373,7 +373,7 @@ class HFLM(LM):
                 assert len(continuation_enc) > 0
                 assert len(continuation_enc) <= self.max_length
 
-                # how this all works:
+                # how this all works (illustrated on a causal decoder-only setup):
                 #          CTX      CONT
                 # inp    0 1 2 3|4 5 6 7 8 9   <- last token is deleted by inp[:, :-1]
                 # model  \               \
@@ -519,7 +519,6 @@ class HFLM(LM):
                 max_gen_toks = gen_kwargs.pop("max_gen_toks")
             else:
                 max_gen_toks = self.max_gen_toks
-
             # first stop sequence is used to halt generation upon encountering
             (primary_until) = until[0]
 
@@ -552,7 +551,8 @@ class HFLM(LM):
 
             # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
             for term in until:
-                s = s.split(term)[0]
+                if len(term) > 0:  # ignore '' separator, for seq2seq case where
+                    s = s.split(term)[0]
 
             res.append(s)