[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
+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