Overview

Preserve each Qwen3-VL video's effective sampling options across repeated items and partial multimodal-cache hits. Reject option lists that are too short before indexing them.

Claims

• Each video's selected fps and num_frames contribute to its cache identity.
• Partial cache hits filter option lists together with the remaining videos.
• Configured defaults, request options, flat/nested scopes, UUID reuse, and encoder profiling preserve their existing behavior.
• Short option lists raise a clear ValueError with and without a processor cache. Unused trailing entries remain supported.

Validation

Validated on 2026-10-11 at head a0067713d3, rebased onto upstream 01abe975b5.

VLLM_TARGET_DEVICE=cpu CUDA_VISIBLE_DEVICES='' uv run --no-project -- .venv/bin/python -m pytest tests/models/multimodal/processing/test_qwen3_vl.py -q
VLLM_TARGET_DEVICE=cpu CUDA_VISIBLE_DEVICES='' uv run --no-project -- .venv/bin/python -m pytest tests/multimodal/test_cache.py tests/multimodal/test_processing.py -q
uv run --no-project -- .venv/bin/pre-commit run --hook-stage manual --files vllm/multimodal/processing/processor.py vllm/model_executor/models/qwen3_vl.py tests/models/multimodal/processing/test_qwen3_vl.py

The Qwen3 processor, generic processing, and cache suites pass, as do the applicable scoped hooks. The CUDA sleep/wake integration was skipped in the CPU suite. The short-list reproducer failed with IndexError: list index out of range before the guard and raises the expected ValueError afterward. Regressions exercise empty/short lists, warm caches, longer lists through partial hits, and profiling with fewer videos than configured options.

The cache regression checks prompts, placeholders, tensors, hashes, scalar-reference grids, duplicate videos, partial hits, reordering, model/request options, flat/nested scopes, and UUID-only reuse. Input options remain unmodified.

Validation used the shared development environment, pinned Humming 0.1.16, and existing precompiled CUDA extensions. Optional FlashInfer autotuning and JIT warmup were disabled. This is a focused 2B family comparison, not a 4B answer-quality benchmark.

Details

Request-wide option-list hashes let repeated videos reuse the first video's result despite different sampling settings. Partial hits removed videos without removing the corresponding option entries.

The existing base hooks let Qwen3-VL refine item hashes and select cache-miss options while preserving other processors. The final text/image HF call suppresses configured video lists so merging cannot reintroduce them. The review fix stays in the Qwen3 processor and its existing regression suite.

Merged #56372 handles option scoping and the cache API; this fix uses that behavior. The current overlap search found no competing per-item cache identity/partial-hit fix. #54527 addresses scoped option reads in other models; #60879 bounds MiniMax video decoding.

Affected hashes retain the full option list as well as selected item options, so equivalent reordered requests can incur extra misses. Scalar paths retain their existing hashes.

Upstream pre-commit is blocked at contributor authorization before implementation checks run. Read the Docs reports success from skipping this change because no docs-affecting files changed; MkDocs did not run. AI assistance was used for investigation, implementation, validation, and review.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)