From f49786ccba1ab17c78ba880b9d5efe4fd715e5a2 Mon Sep 17 00:00:00 2001
From: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com>
Date: Thu, 9 Mar 2023 11:33:28 +0100
Subject: [PATCH] fix(server): do not warp prefill logits

---
 server/text_generation_server/utils/tokens.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/server/text_generation_server/utils/tokens.py b/server/text_generation_server/utils/tokens.py
index aa76b6eb..15c1b13b 100644
--- a/server/text_generation_server/utils/tokens.py
+++ b/server/text_generation_server/utils/tokens.py
@@ -70,7 +70,11 @@ 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)