Use server-generated keys for late-interaction query caches (#51445)

Co-authored-by: Lucas Bourtoule <35483370+dhalf@users.noreply.github.com>
This commit is contained in:
Clinton Thomas
2026-09-03 07:15:58 +00:00
committed by GitHub
co-authored by Lucas Bourtoule
parent 3f1af35e01
commit ee17d0d869
3 changed files with 75 additions and 2 deletions
@@ -0,0 +1,65 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import AsyncMock, Mock
import pytest
from vllm.entrypoints.pooling.scoring.serving import ServingScores
from vllm.entrypoints.pooling.typing import PoolingServeContext
from vllm.inputs import tokens_input
from vllm.pooling_params import PoolingParams
def _make_context() -> PoolingServeContext:
return PoolingServeContext(
request=Mock(),
model_name="model",
request_id="score-caller-controlled-id",
pooling_params=PoolingParams(task="token_embed"),
lora_request=None,
priorities=None,
prompt_extras=None,
engine_inputs=[
tokens_input(prompt_token_ids=[1]),
tokens_input(prompt_token_ids=[2]),
],
n_queries=1,
)
def _query_key(context: PoolingServeContext) -> str:
assert context.engine_inputs is not None
pooling_params = context.engine_inputs[0]["params"]
late_interaction_params = pooling_params.late_interaction_params
assert late_interaction_params is not None
query_key = late_interaction_params.query_key
assert query_key is not None
return query_key
@pytest.mark.asyncio
async def test_colliding_request_ids_use_distinct_query_cache_keys():
serving = object.__new__(ServingScores)
serving._prepare_generators = AsyncMock()
serving._collect_batch = AsyncMock()
first = _make_context()
second = _make_context()
for context in (first, second):
await serving._flash_late_interaction_encode_queries(context)
await serving._flash_late_interaction_encode_docs(context)
prepared_contexts = [
call.args[0] for call in serving._prepare_generators.await_args_list
]
first_query_key = _query_key(prepared_contexts[0])
first_doc_key = _query_key(prepared_contexts[1])
second_query_key = _query_key(prepared_contexts[2])
second_doc_key = _query_key(prepared_contexts[3])
assert first_query_key == first_doc_key
assert second_query_key == second_doc_key
assert first_query_key != second_query_key
assert "caller-controlled-id" not in first_query_key
assert "caller-controlled-id" not in second_query_key
+9 -2
View File
@@ -9,6 +9,7 @@ from vllm.entrypoints.serve.engine.protocol import UsageInfo
from vllm.logger import init_logger
from vllm.outputs import PoolingRequestOutput, ScoringRequestOutput
from vllm.tasks import SCORE_TYPE_MAP, SupportedTask
from vllm.utils import random_uuid
from vllm.v1.pool.late_interaction import (
build_late_interaction_doc_params,
build_late_interaction_query_params,
@@ -208,7 +209,11 @@ class ServingScores(PoolingServing):
n_docs = len(ctx.engine_inputs) - n_queries
query_engine_inputs = ctx.engine_inputs[:n_queries]
query_keys = [f"{ctx.request_id}-query-{i}" for i in range(n_queries)]
query_namespace = random_uuid()
query_keys = [
f"late-interaction-{query_namespace}-query-{i}" for i in range(n_queries)
]
ctx.late_interaction_query_keys = query_keys
query_uses = [n_docs if n_queries == 1 else 1] * n_queries
for i in range(n_queries):
@@ -249,7 +254,9 @@ class ServingScores(PoolingServing):
n_docs = len(ctx.engine_inputs) - n_queries
doc_engine_inputs = ctx.engine_inputs[n_queries:]
query_keys = [f"{ctx.request_id}-query-{i}" for i in range(n_queries)]
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)]
for i in range(n_docs):
+1
View File
@@ -114,6 +114,7 @@ class PoolingServeContext(Generic[PoolingRequestT]):
## for flash-late-interaction
query_final_res_batch: list[PoolingRequestOutput] | None = None
late_interaction_query_keys: list[str] | None = None
@dataclass