[BugFix] Fix P/D with non-MoE DP (#33037)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-01-27 08:03:47 -08:00
committed by GitHub
parent 83fb2d09e8
commit 0cd259b2d8
2 changed files with 18 additions and 11 deletions

View File

@@ -313,6 +313,13 @@ class CoreEngineActorManager:
dp_vllm_config.parallel_config.placement_group = pg
local_client = index < local_engine_count
if dp_size > 1 and dp_vllm_config.kv_transfer_config is not None:
# modify the engine_id and append the local_dp_rank to it to ensure
# that the kv_transfer_config is unique for each DP rank.
dp_vllm_config.kv_transfer_config.engine_id = (
f"{dp_vllm_config.kv_transfer_config.engine_id}_dp{local_index}"
)
# Ray XPU known issue: dpctl initializes the GPU runtime early, so
# setting device env vars in Ray actor's initialization method
# will not affect device selection. See: