[BowenFu]

Require aiohttp~=3.14.4 and pillow~=13.0.0 for their security fixes. Preserve existing custom-image behavior: Pillow 13 rejects 32-bit integer PNG encoding, so convert mode I to I;16 with the same clipping Pillow 12 used. Python and minor-series bounds remain unchanged.

Validation:

• Three focused cases cover integer pixels, input immutability and real TIFF resize/PNG/JPEG encoding. Both PNG cases fail before the fix and pass afterward; the original Pillow 12.2 control passes.
• Four PNG outputs match the original Pillow 12.2 bytes and pixels exactly.
• Dataset/plot coverage: 3,785 passed across the initial run and unchanged setup retries; five skips and one expected failure remain. The 24 initial failures came from the disposable container's missing user home.
• Pre-commit passed. Prior full dependency validation retains its two documented baseline failures.

No performance claim.

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@779feae3741d639e5cdcaafb1e99f41cda15c8c6

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@779feae3741d639e5cdcaafb1e99f41cda15c8c6

---

Last updated for commit: 779feae • Browse code

[BowenFu]

Withdrawn. The TensorRT-LLM consumer integration is tracked in NVIDIA/TensorRT-LLM#21322. I am closing this upstream PR to keep the work scoped to TensorRT-LLM. The source commits remain available.