[warmup] overlap renderer warmup and engine core initialization (#54557)

Signed-off-by: Andy Xie <andy.xning@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Ning Xie
2026-09-03 19:44:21 +00:00
committed by GitHub
co-authored by Cyrus Leung
parent fc8f10792c
commit 2a336d8239
5 changed files with 561 additions and 33 deletions
+8
View File
@@ -102,12 +102,17 @@ class LLMEngine:
)
# EngineCore (gets EngineCoreRequests and gives EngineCoreOutputs)
# Hand the renderer to the client. In multiprocess mode the client
# starts the MM warmup only after engine-core fork (the why is in
# BaseRenderer.start_mm_warmup_in_background); InprocClient takes no
# renderer, so MM warmup stays inside renderer.warmup() there.
self.engine_core = EngineCoreClient.make_client(
multiprocess_mode=multiprocess_mode,
asyncio_mode=False,
vllm_config=vllm_config,
executor_class=executor_class,
log_stats=self.log_stats,
renderer=renderer,
)
self.logger_manager: StatLoggerManager | None = None
@@ -343,6 +348,9 @@ class LLMEngine:
self.engine_core.profile(False)
def reset_mm_cache(self):
# Join the background MM warmup first: the mm_processor_cache is not
# safe for concurrent access with its apply/clear.
self.renderer._join_mm_warmup()
self.renderer.clear_mm_cache()
self.engine_core.reset_mm_cache()