[Bugfix][Model Runner V2] Fix min_tokens off-by-one in the V2 GPU sampler (#46243)
Signed-off-by: Ting Sun <suntcrick@gmail.com>
Ting SUN committed
Jun 21, 2026 at 12:06 UTC
183a430c137db3d5cd0b9025b816f26ee87328e7
1 file changed
+1
-1
vllm/v1/worker/gpu/sample/logit_bias.py
+1
-1
@@ -222,7 +222,7 @@ def _bias_kernel(
222
num_stop_token_ids = tl.load(num_stop_token_ids_ptr + req_state_idx)
223
pos = tl.load(pos_ptr + token_idx)
224
min_len = tl.load(min_lens_ptr + req_state_idx)
225
- if num_stop_token_ids > 0 and pos < min_len:
225
+ if num_stop_token_ids > 0 and pos + 1 < min_len:
226
mask = block < num_stop_token_ids
227
stop_token_ids = tl.load(
228
stop_token_ids_ptr + req_state_idx * stop_token_ids_stride + block,