[Bugfix][Pooling] Handle multimodal cache misses without crashing (#58975)

Signed-off-by: linchaoying6 <linchaoying6@jd.com>
Co-authored-by: linchaoying6 <linchaoying6@jd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Chaoying
2026-10-02 10:14:33 -04:00
committed by GitHub
co-authored by linchaoying6 OpenAI Codex
parent 443dccd3c9
commit 592c6f3fbc
21 changed files with 525 additions and 17 deletions
+30 -5
View File
@@ -195,13 +195,36 @@ class ServingScores(PoolingServing):
ctx = await self._init_ctx(self.io_processor, *args, **kwargs)
await self._preprocessing(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)
try:
# 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)
except BaseException:
try:
await self._cleanup_flash_late_interaction(ctx)
except Exception:
logger.exception("Failed to clean up late-interaction query cache.")
raise
return await self._postprocessing_async(self.io_processor, ctx)
async def _cleanup_flash_late_interaction(self, ctx: ScoringServeContext) -> None:
query_keys = ctx.late_interaction_query_keys
if not query_keys:
return
doc_keys = ctx.late_interaction_doc_keys or []
try:
# Stop documents before removing the query tensors they reference.
await self.engine_client.abort([*query_keys, *doc_keys])
finally:
await self.engine_client.collective_rpc(
"release_late_interaction_query_cache",
args=(query_keys,),
)
async def _flash_late_interaction_encode_queries(self, ctx: ScoringServeContext):
assert ctx.n_queries is not None
assert ctx.engine_inputs is not None
@@ -259,7 +282,9 @@ class ServingScores(PoolingServing):
query_keys = ctx.late_interaction_query_keys
if query_keys is None:
raise RuntimeError("Late-interaction query keys were not initialized.")
doc_keys = [f"{ctx.request_id}-doc-{i}" for i in range(n_docs)]
doc_namespace = random_uuid()
doc_keys = [f"late-interaction-{doc_namespace}-doc-{i}" for i in range(n_docs)]
ctx.late_interaction_doc_keys = doc_keys
for i in range(n_docs):
query_idx = 0 if n_queries == 1 else i