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 13, 19:30 UTC
[bugfix] fix wrong `dcp_local_seq_lens` calc (#27518)
Signed-off-by: Qiu <qiuchunshuo@huawei.com>
Qiu committed
Nov 5, 2025 at 17:58 UTC
16b37f3119918c1e5a39f303e0d0892c65c07a90
1 file changed
+1
-1
vllm/v1/attention/backends/mla/common.py
+1
-1
@@ -800,7 +800,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
800
dcp_local_seq_lens[:num_decodes] = seq_lens[
801
:num_decodes
802
] // self.dcp_world_size + (
803
- self.dcp_rank <= (seq_lens[:num_decodes] - 1) % self.dcp_world_size
803
+ self.dcp_rank < seq_lens[:num_decodes] % self.dcp_world_size
804
)
805
806
assert num_decodes + num_prefills == num_reqs