fix(server): do not warp prefill logits (#116)

server/text_generation_server/utils/tokens.py

@@ -75,6 +75,10 @@ class NextTokenChooser:
 
     def __call__(self, input_ids, scores):
         # Warp logits
+        if scores.shape[0] > 1:
+            # only warp the last token logits
+            scores[-1:, :] = self.warpers(input_ids, scores[-1:, :])
+        else:
             scores = self.warpers(input_ids, scores)
 
         # Compute logprobs