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 8, 15:15 UTC
[Bugfix] Fix DP Attention Padding in Dummy Run (#34187)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>
Lucas Wilkinson committed
Feb 9, 2026 at 21:29 UTC
81e217fe6b5a3030aa5c4d859a2125b81979bee4
1 file changed
+1
vllm/v1/worker/gpu_model_runner.py
+1
@@ -4787,6 +4787,7 @@ class GPUModelRunner(
4787
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
4788
attn_metadata, _ = self._build_attention_metadata(
4789
num_tokens=num_tokens_unpadded,
4790
+ num_tokens_padded=num_tokens_padded if pad_attn else None,
4791
num_reqs=num_reqs_padded,
4792
max_query_len=max_query_len,
4793
ubatch_slices=ubatch_slices_padded if pad_attn else ubatch_slices,