wrap self._model_call with F.log_softmax + remove empty_cache, not needed

lm_eval/base.py

@@ -197,7 +197,7 @@ class BaseLM(LM):
             @find_executable_batch_size(starting_batch_size=512) # if OOM, then halves batch_size and tries again
             def forward_batch(batch_size):
                 test_batch = torch.ones((batch_size, self.max_length), device=self.device).long()
-                self._model_call(test_batch) 
+                F.log_softmax(self._model_call(test_batch), dim = -1)
                 return batch_size
             
             batch_size = forward_batch() 
@@ -264,7 +264,7 @@ class BaseLM(LM):
                 @find_executable_batch_size(starting_batch_size=512) # if OOM, then halves batch_size and tries again
                 def forward_batch(batch_size):
                     test_batch = torch.ones((batch_size, max_context), device=self.device).long()
-                    self._model_call(test_batch) 
+                    F.log_softmax(self._model_call(test_batch), dim = -1)
                     return batch_size
                 
                 batch_size = forward_batch() 
@@ -274,8 +274,6 @@ class BaseLM(LM):
             else:
                 adaptive_batch_size = override_bs
         
-        torch.cuda.empty_cache() # empty cache after determining batch size
-
         for chunk in utils.chunks(
             tqdm(re_ord.get_reordered(), disable=disable_tqdm), self.batch_size if self.batch_size != "auto" else adaptive_batch_size
         ):