[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
@@ -1,20 +1,25 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for ScoringIOProcessor post-tokenization helpers."""
"""Unit tests for ScoringIOProcessor helpers."""
from dataclasses import dataclass
from types import SimpleNamespace
import pytest
import torch
from tokenizers import Tokenizer, models, pre_tokenizers, processors
from transformers import TokenizersBackend
from vllm import TokensPrompt
from vllm.entrypoints.pooling.scoring.io_processor import (
BiEncoderIOProcessor,
CrossEncoderIOProcessor,
JinaRankingIOProcessor,
LateInteractionIOProcessor,
_apply_post_tokenization_to_token_type_ids,
)
from vllm.entrypoints.pooling.scoring.utils import compress_token_type_ids
from vllm.outputs import PoolingOutput, PoolingRequestOutput, RequestError
from vllm.renderers import TokenizeParams
pytestmark = pytest.mark.skip_global_cleanup
@@ -113,3 +118,59 @@ def test_token_type_ids_stay_aligned_with_a_truncated_padded_prompt():
assert first_doc == 10
assert prompt_token_ids[:first_doc] == list(range(10, num_query))
assert prompt_token_ids[first_doc:40] == list(range(num_query, 50))
def _pooling_output(
request_id: str,
data: torch.Tensor,
error: RequestError | None = None,
) -> PoolingRequestOutput:
return PoolingRequestOutput(
request_id=request_id,
outputs=PoolingOutput(data),
prompt_token_ids=[1],
num_cached_tokens=0,
finished=True,
error=error,
)
@pytest.mark.parametrize(
"processor_type", [BiEncoderIOProcessor, LateInteractionIOProcessor]
)
@pytest.mark.parametrize("failed_index", [0, 1])
def test_paired_scoring_preserves_request_error(processor_type, failed_index):
processor = processor_type.__new__(processor_type)
processor.pad_token_id = None
error = RequestError(
code="multimodal_cache_miss",
message="Multi-modal processor cache miss.",
retryable=True,
)
outputs = [
_pooling_output("query", torch.tensor([1.0, 2.0])),
_pooling_output("document", torch.tensor([3.0, 4.0])),
]
outputs[failed_index] = _pooling_output(
outputs[failed_index].request_id,
torch.empty(0),
error,
)
(result,) = processor._post_process(outputs, n_queries=1)
assert result.request_id == "query_document"
assert result.error is error
assert result.outputs is outputs[failed_index].outputs
def test_jina_scoring_preserves_request_error_before_indexing_output():
processor = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)
error = RequestError(
code="multimodal_cache_miss",
message="Multi-modal processor cache miss.",
retryable=True,
)
failed = _pooling_output("request", torch.empty(0), error)
assert processor._post_process([failed], n_queries=1) == [failed]
@@ -49,10 +49,60 @@ def _query_key(context: PoolingServeContext) -> str:
@pytest.mark.asyncio
async def test_colliding_request_ids_use_distinct_query_cache_keys(
@pytest.mark.skip_global_cleanup
@pytest.mark.parametrize("failed_stage", ["query", "doc"])
async def test_flash_late_interaction_cleans_up_queries_on_failure(
failed_stage: str,
monkeypatch: pytest.MonkeyPatch,
):
serving = object.__new__(ServingScores)
serving.io_processor = Mock()
serving.engine_client = Mock(abort=AsyncMock(), collective_rpc=AsyncMock())
context = _make_context()
monkeypatch.setattr(serving, "_init_ctx", AsyncMock(return_value=context))
monkeypatch.setattr(serving, "_preprocessing", AsyncMock())
async def encode_queries(ctx):
ctx.late_interaction_query_keys = ["query-key"]
if failed_stage == "query":
raise RuntimeError("query failed")
async def encode_docs(ctx):
ctx.late_interaction_doc_keys = ["doc-key"]
if failed_stage == "doc":
raise RuntimeError("doc failed")
monkeypatch.setattr(
serving,
"_flash_late_interaction_encode_queries",
AsyncMock(side_effect=encode_queries),
)
monkeypatch.setattr(
serving,
"_flash_late_interaction_encode_docs",
AsyncMock(side_effect=encode_docs),
)
with pytest.raises(RuntimeError, match=f"{failed_stage} failed"):
await serving.flash_late_interaction()
expected_doc_keys = [] if failed_stage == "query" else ["doc-key"]
serving.engine_client.abort.assert_awaited_once_with(
["query-key", *expected_doc_keys]
)
serving.engine_client.collective_rpc.assert_awaited_once_with(
"release_late_interaction_query_cache",
args=(["query-key"],),
)
@pytest.mark.asyncio
async def test_colliding_request_ids_use_distinct_late_interaction_keys(
monkeypatch: pytest.MonkeyPatch,
):
serving = object.__new__(ServingScores)
serving.engine_client = Mock(abort=AsyncMock(), collective_rpc=AsyncMock())
prepare_generators = AsyncMock()
monkeypatch.setattr(serving, "_prepare_generators", prepare_generators)
monkeypatch.setattr(serving, "_collect_batch", AsyncMock())
@@ -68,9 +118,24 @@ async def test_colliding_request_ids_use_distinct_query_cache_keys(
first_doc_key = _query_key(prepared_contexts[1])
second_query_key = _query_key(prepared_contexts[2])
second_doc_key = _query_key(prepared_contexts[3])
first_doc_request_ids = prepared_contexts[1].prompt_request_ids
second_doc_request_ids = prepared_contexts[3].prompt_request_ids
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
assert first_doc_request_ids is not None
assert second_doc_request_ids is not None
assert first.late_interaction_doc_keys == first_doc_request_ids
assert second.late_interaction_doc_keys == second_doc_request_ids
assert first_doc_request_ids != second_doc_request_ids
assert all("caller-controlled-id" not in key for key in first_doc_request_ids)
assert all("caller-controlled-id" not in key for key in second_doc_request_ids)
assert first.late_interaction_query_keys is not None
await serving._cleanup_flash_late_interaction(first)
serving.engine_client.abort.assert_awaited_once_with(
[*first.late_interaction_query_keys, *first_doc_request_ids]
)
+43 -1
View File
@@ -2,8 +2,17 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from vllm.outputs import RequestOutput
from vllm.outputs import (
ClassificationRequestOutput,
EmbeddingRequestOutput,
PoolingOutput,
PoolingRequestOutput,
RequestError,
RequestOutput,
ScoringRequestOutput,
)
pytestmark = pytest.mark.cpu_test
@@ -19,3 +28,36 @@ def test_request_output_forward_compatible():
example_arg_added_in_new_version="some_value",
)
assert output is not None
@pytest.mark.parametrize(
("output_type", "empty_value"),
[
(EmbeddingRequestOutput, []),
(ClassificationRequestOutput, []),
(ScoringRequestOutput, 0.0),
],
)
@pytest.mark.skip_global_cleanup
def test_specialized_pooling_outputs_preserve_request_error(output_type, empty_value):
error = RequestError(
code="multimodal_cache_miss",
message="Multi-modal processor cache miss.",
retryable=True,
)
base_output = PoolingRequestOutput(
request_id="request",
outputs=PoolingOutput(torch.empty(0)),
prompt_token_ids=[1, 2],
num_cached_tokens=0,
finished=True,
error=error,
)
output = output_type.from_base(base_output)
assert output.error is error
value = getattr(output.outputs, "embedding", None)
value = getattr(output.outputs, "probs", value)
value = getattr(output.outputs, "score", value)
assert value == empty_value
+34
View File
@@ -2,11 +2,13 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import random
from typing import TYPE_CHECKING
from unittest.mock import MagicMock, call
import pytest
from vllm import LLM
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
from vllm.v1.engine.llm_engine import LLMEngine
from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Metric, Vector
if TYPE_CHECKING:
@@ -18,6 +20,38 @@ MODEL = "facebook/opt-125m"
DTYPE = "half"
@pytest.mark.skip_global_cleanup
def test_step_invalidates_mm_processor_cache_misses():
engine = object.__new__(LLMEngine)
engine.should_execute_dummy_batch = False
engine.log_stats = False
engine.logger_manager = None
mm_processor_cache = MagicMock()
engine.renderer = MagicMock(mm_processor_cache=mm_processor_cache)
engine_core_output = MagicMock(
mm_cache_miss_hashes=["missing-hash-1", "missing-hash-2"]
)
outputs = MagicMock(
outputs=[engine_core_output],
timestamp=0.0,
scheduler_stats=None,
)
engine.engine_core = MagicMock()
engine.engine_core.get_output.return_value = outputs
processed_outputs = MagicMock(reqs_to_abort=[], request_outputs=[])
engine.output_processor = MagicMock()
engine.output_processor.process_outputs.return_value = processed_outputs
assert engine.step() == []
assert mm_processor_cache.invalidate.call_args_list == [
call("missing-hash-1"),
call("missing-hash-2"),
]
def _vllm_model(
apc: bool,
vllm_runner: type[VllmRunner],
+96 -1
View File
@@ -7,6 +7,7 @@ from unittest.mock import MagicMock, Mock
import numpy as np
import pytest
import torch
from tests.v1.engine.utils import (
NUM_PROMPT_LOGPROBS_UNDER_TEST,
@@ -18,7 +19,12 @@ from tests.v1.engine.utils import (
from vllm import PoolingParams
from vllm.logprobs import FlatLogprobs, Logprob, PromptLogprobs, SampleLogprobs
from vllm.lora.request import LoRARequest
from vllm.outputs import CompletionOutput, RequestOutput, SamplingMask
from vllm.outputs import (
CompletionOutput,
PoolingRequestOutput,
RequestOutput,
SamplingMask,
)
from vllm.sampling_params import RequestOutputKind, SamplingParams
from vllm.tokenizers import TokenizerLike
from vllm.v1.engine import (
@@ -1630,3 +1636,92 @@ def test_request_output_add_merges_delta_sampling_masks():
assert merged.outputs[0].token_ids == [10, 20]
assert merged.outputs[0].sampling_mask.token_ids == [[10, 11], [20]]
@pytest.mark.skip_global_cleanup
def test_pooling_cache_miss_is_a_request_error_and_batch_continues():
output_processor = OutputProcessor(tokenizer=None, log_stats=False)
requests = [
EngineCoreRequest(
request_id=f"request-{idx}-int",
external_req_id=f"request-{idx}",
prompt_token_ids=[1, 2],
mm_features=None,
arrival_time=0,
lora_request=None,
cache_salt=None,
data_parallel_rank=None,
sampling_params=None,
pooling_params=PoolingParams(task="embed"),
)
for idx in range(2)
]
for request in requests:
output_processor.add_request(request, prompt=None)
processed = output_processor.process_outputs(
[
EngineCoreOutput(
request_id=requests[0].request_id,
new_token_ids=[],
finish_reason=FinishReason.ERROR,
mm_cache_miss_hashes=["missing-hash"],
),
EngineCoreOutput(
request_id=requests[1].request_id,
new_token_ids=[],
pooling_output=torch.tensor([1.0, 2.0]),
finish_reason=FinishReason.STOP,
),
]
)
assert len(processed.request_outputs) == 2
failed, succeeded = processed.request_outputs
assert isinstance(failed, PoolingRequestOutput)
assert failed.error is not None
assert failed.error.code == "multimodal_cache_miss"
assert failed.error.retryable
assert isinstance(succeeded, PoolingRequestOutput)
assert succeeded.error is None
assert torch.equal(succeeded.outputs.data, torch.tensor([1.0, 2.0]))
@pytest.mark.skip_global_cleanup
def test_async_pooling_cache_miss_does_not_fail_output_processor():
output_processor = OutputProcessor(tokenizer=None, log_stats=False)
request = EngineCoreRequest(
request_id="request-int",
external_req_id="request",
prompt_token_ids=[1, 2],
mm_features=None,
arrival_time=0,
lora_request=None,
cache_salt=None,
data_parallel_rank=None,
sampling_params=None,
pooling_params=PoolingParams(task="embed"),
)
queue = RequestOutputCollector(
output_kind=RequestOutputKind.FINAL_ONLY,
request_id=request.request_id,
)
output_processor.add_request(request, prompt=None, queue=queue)
processed = output_processor.process_outputs(
[
EngineCoreOutput(
request_id=request.request_id,
new_token_ids=[],
finish_reason=FinishReason.ERROR,
mm_cache_miss_hashes=["missing-hash"],
)
]
)
assert processed.request_outputs == []
output = queue.get_nowait()
assert isinstance(output, PoolingRequestOutput)
assert output.error is not None
assert output.error.retryable
assert not output_processor.has_unfinished_requests()
@@ -137,6 +137,52 @@ def test_finished_request_releases_unscored_doc_use():
)
@pytest.mark.skip_global_cleanup
def test_release_queries_is_targeted_and_idempotent():
runner = LateInteractionRunner()
query_emb = torch.tensor([[1.0, 0.0], [0.0, 1.0]], dtype=torch.float32)
for query_key in ("query-release", "query-keep"):
query_params = _make_pooling_params(
build_late_interaction_query_params(query_key=query_key, query_uses=1)
)
runner.postprocess_pooler_output(
raw_pooler_output=[query_emb],
pooling_params=[query_params],
req_ids=[f"{query_key}-req"],
finished_mask=[True],
)
released_doc_params = _make_pooling_params(
build_late_interaction_doc_params(query_key="query-release")
)
kept_doc_params = _make_pooling_params(
build_late_interaction_doc_params(query_key="query-keep")
)
runner.register_request("released-doc", released_doc_params)
runner.register_request("kept-doc", kept_doc_params)
runner.release_queries(["query-release"])
runner.release_queries(["query-release"])
with pytest.raises(ValueError, match="query cache miss"):
runner.postprocess_pooler_output(
raw_pooler_output=[query_emb],
pooling_params=[released_doc_params],
req_ids=["released-doc"],
finished_mask=[True],
)
kept_output = runner.postprocess_pooler_output(
raw_pooler_output=[query_emb],
pooling_params=[kept_doc_params],
req_ids=["kept-doc"],
finished_mask=[True],
)
assert isinstance(kept_output, list)
assert kept_output[0] is not None
def test_invalid_query_uses_raises():
runner = LateInteractionRunner()
bad_meta = LateInteractionParams(