[None][perf] Improve Qwen3-VL Preprocessing Perf#15598

Open

aswinvisva wants to merge 19 commits into

NVIDIA:mainfrom

aswinvisva:avisva/cosmos3-vlm-preproc-perf

Commits on Jul 1, 2026

serve: install uvloop as default asyncio event loop policy
aswinvisva
committed
inputs/media_io: route media-load work through a dedicated ThreadPoolExecutor
aswinvisva
committed
serve/openai_server: route HF input-processor calls through a dedicated ThreadPoolExecutor
aswinvisva
committed
inputs/media_io: BytesIO + uint8 numpy fast path for video decode
aswinvisva
committed
qwen3vl: skip HF frame sampling when frames are already sampled by the IO loader
aswinvisva
committed
qwen2/3vl: memoize HF processor _merge_kwargs by call signature
aswinvisva
committed
requirements: add uvloop
aswinvisva
committed
address review feedback
aswinvisva
committed
serve: install uvloop after all imports to avoid try/except wrapper
aswinvisva
committed
address review feedback
aswinvisva
committed
qwen3vl: strip io_loaded_all_frames before HF VideoMetadata
aswinvisva
committed
swap signature-keyed merge_kwargs memoize for HF schema cache
aswinvisva
committed
media_io: lazy_loader for cv2 + unit tests for new control flow
aswinvisva
committed
address review feedback: media_io rename, comment, staticmethod; drop io_loaded_all_frames stash
aswinvisva
committed
media_io: suppress bandit B108 on /dev/shm tempdir literal
aswinvisva
committed
qwen3vl: null fps when caller sets only num_frames so HF respects the count
aswinvisva
committed
qwen3vl: sample video frames when caller gives no num_frames/fps
aswinvisva
committed
qwen3vl: update silent-caller do_sample_frames unit test
aswinvisva
committed

Commits on Jul 2, 2026

qwen3vl: register new serve CLI options in API stability reference
aswinvisva
committed