[Frontend] Offload blocking preprocessing & postprocessing ops to thread pool for pooling entrypoints. (#39763)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
wang.yuqi
2026-04-14 08:29:25 +00:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent 0008729abf
commit c0ecaed950
10 changed files with 68 additions and 83 deletions
+4 -6
View File
@@ -65,7 +65,7 @@ class ServingScores(PoolingServing):
return await self.flash_late_interaction(*args, **kwargs)
async def _build_response(
def _build_response(
self,
ctx: ScoringServeContext,
) -> JSONResponse:
@@ -183,17 +183,15 @@ class ServingScores(PoolingServing):
### Can significantly improve late-interaction scoring performance.
async def flash_late_interaction(self, *args, **kwargs) -> Response:
ctx = await self._init_ctx(*args, **kwargs)
ctx.pooling_params = self.io_processor.create_pooling_params(ctx.request)
await self.io_processor.pre_process_online_async(ctx)
ctx = await self._init_ctx(self.io_processor, *args, **kwargs)
await self._preprocessing_async(self.io_processor, ctx)
# stage 1: encode queries and cache token embeddings on workers.
await self._flash_late_interaction_encode_queries(ctx)
# stage 2: encode docs and return scalar scores from workers.
await self._flash_late_interaction_encode_docs(ctx)
await self.io_processor.post_process_online_async(ctx)
return await self._build_response(ctx)
return await self._postprocessing_async(self.io_processor, ctx)
async def _flash_late_interaction_encode_queries(self, ctx: ScoringServeContext):
assert ctx.n_queries is not None