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 11, 23:45 UTC
[Bugfix] Correct num_q_heads on DCP for Flashinfer backends (#29487)
Signed-off-by: Jingchun Gao <gaojingchun1@huawei.com> Signed-off-by: Jingchun Gao <63247409+gjc0824@users.noreply.github.com> Co-authored-by: Jingchun Gao <gaojingchun1@huawei.com>
Jingchun Gao committed
Dec 5, 2025 at 13:54 UTC
d698bb382db95af6b8836936eb0dfae71c791d11
1 file changed
+2
-3
vllm/v1/attention/backends/flashinfer.py
+2
-3
@@ -482,9 +482,8 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
482
self.dcp_rank = 0
483
self.dcp_kv_cache_interleave_size = 1
484
485
- self.num_qo_heads = (
486
- self.model_config.get_num_attention_heads(self.vllm_config.parallel_config)
487
- * self.dcp_world_size
485
+ self.num_qo_heads = self.model_config.get_num_attention_heads(
486
+ self.vllm_config.parallel_config
487
)
488
489
self.num_kv_heads = self.kv_cache_spec.num_kv_heads