A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 04:30 UTC
[Bugfix] RoBERTa position_id accumulation in CUDA graph padding region (#37873)
Signed-off-by: dass90 <3053034939@qq.com>
yanghui1-arch committed
Mar 23, 2026 at 22:59 UTC
7151ae6528805ad474745c10516a328a36978de9
1 file changed
+2
vllm/v1/worker/gpu_model_runner.py
+2
@@ -3084,6 +3084,8 @@ class GPUModelRunner(
3084
positions = self.xdrope_positions.gpu[:, :num_input_tokens]
3085
else:
3086
positions = self.positions.gpu[:num_input_tokens]
3087
+ if num_input_tokens > num_scheduled_tokens:
3088
+ self.positions.gpu[num_scheduled_tokens:num_input_tokens].zero_()
3089
3090
if is_first_rank:
3091
intermediate_tensors = None