Offering Refactoring Expertise for Multimodal Bugs (#8652) – 20+ Yrs Exp in AI/ML

[pythonjavaerlang]

NVIDIA, your TensorRT-LLM 1.2.0rc0 multimodal runner is unusable in production:

• IndexError on batch image inputs (always returns 1 image despite image_path list)
• Undefined sampling_config_list when sampling_config=None
• In-method imports, duplicated SamplingConfig defs (Python vs C++), ignored CLI args

I debugged this for a week on an L4 GPU (€1600) — nothing works reliably.

With 20+ years in Python/Java/Erlang/AI, I can refactor this cleanly in days.
Let’s fix it before it damages trust in NVIDIA’s open-source stack.

GitHub Issue: #8652
LinkedIn: https://www.linkedin.com/in/python-java-erlang-ai-ml-developer/