[Bugfix] Sync block_size from EngineCore to frontend for hybrid Mamba… (#42967)

Signed-off-by: Amit Gruner <agruner@crusoe.ai>
Co-authored-by: Amit Gruner <agruner@crusoe.ai>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
This commit is contained in:
gruner
2026-06-02 13:41:00 +00:00
committed by GitHub
co-authored by Amit Gruner Jiangyun Zhu
parent b623f7ea95
commit 654bd2bca4
4 changed files with 32 additions and 1 deletions
+4
View File
@@ -713,6 +713,10 @@ class MPClient(EngineCoreClient):
num_gpu_blocks += response.num_gpu_blocks
vllm_config.cache_config.num_gpu_blocks = num_gpu_blocks
# Sync block_size: may be enlarged by _align_hybrid_block_size in the
# worker for hybrid Mamba models.
vllm_config.cache_config.block_size = response.block_size
# In external DP LB mode, the coordinator address that the
# front-end procs connect to is obtained by each engine via it's
# initial handshake with the rank 0 front-end.