[BugFix][V1] Fix overhead related to bad_words sampling when not in use (#14894)

Signed-off-by: Nick Hill <nhill@redhat.com>
2025-03-16 14:53:34 -07:00
parent f6137adbcb
commit fc1f67715d
3 changed files with 10 additions and 7 deletions
--- a/vllm/v1/worker/gpu_input_batch.py
+++ b/vllm/v1/worker/gpu_input_batch.py
@@ -324,8 +324,9 @@ class InputBatch:
            self.allowed_token_ids_mask_cpu_tensor[req_index][
                sampling_params.allowed_token_ids] = False

-        self.bad_words_token_ids[
-            req_index] = sampling_params.bad_words_token_ids
+        if sampling_params.bad_words_token_ids:
+            self.bad_words_token_ids[
+                req_index] = sampling_params.bad_words_token_ids

        # Add request lora ID
        if request.lora_request: