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

c0795de2 · OlivierDehaene · GitHub · 1a2d6825 · c0795de2
Unverified Commit c0795de2 authored Mar 09, 2023 by OlivierDehaene Committed by GitHub Mar 09, 2023
Hide whitespace changes
Inline Side-by-side

Showing with 5 additions and 1 deletion

server/text_generation_server/utils/tokens.py server/text_generation_server/utils/tokens.py +5 -1

No files found.
--- a/server/text_generation_server/utils/tokens.py
+++ b/server/text_generation_server/utils/tokens.py
@@ -75,7 +75,11 @@ class NextTokenChooser:

    def __call__(self, input_ids, scores):
        # Warp logits
-        scores = self.warpers(input_ids, scores)
+        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
        logprobs = torch.log_softmax(scores, -1)