fix p-tuning inaccuracy, because output logit contains virtual token length

Signed-off-by: Wang, Yi <yi.a.wang@intel.com>

lm_eval/base.py

@@ -289,7 +289,6 @@ class BaseLM(LM):
         ):
             inps = []
             cont_toks_list = []
-            inplens = []
 
             padding_length = None
 
@@ -337,19 +336,19 @@ class BaseLM(LM):
 
                 inps.append(inp.unsqueeze(0))  # [1, padding_length]
                 cont_toks_list.append(cont)
-                inplens.append(inplen)
 
             batched_inps = torch.cat(inps, dim=0)  # [batch, padding_length
             multi_logits = F.log_softmax(
                 self._model_call(batched_inps), dim=-1
             ).cpu()  # [batch, padding_length, vocab]
 
-            for (cache_key, _, _), logits, inp, inplen, cont_toks in zip(
-                chunk, multi_logits, inps, inplens, cont_toks_list
+            for (cache_key, _, _), logits, inp, cont_toks in zip(
+                chunk, multi_logits, inps, cont_toks_list
             ):
 
                 # Slice to original seq length
                 contlen = len(cont_toks)
+                inplen = logits.shape[0]
                 logits = logits[inplen - contlen : inplen].unsqueeze(
                     0
                 )  # [1, seq, vocab]