RebelsImageToPrompt crashes ComfyUI process (process hang) with Qwen3.5 mmproj

[AIDAMHF]

Environment

• ComfyUI: 0.22.3
• OS: Windows 11
• GPU: high-end NVIDIA (CUDA 13.0)
• Python: 3.10.11
• PyTorch: 2.12.0+cu130

Models used

• model_file: Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf
• mmproj_file: mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguf
• chat_handler: Auto-detect
• (Both files placed in the node folder as documented in README)

Issue

When executing RebelsImageToPrompt, the entire ComfyUI process crashes and becomes completely unresponsive. It cannot be stopped normally from the launcher and requires a forced kill via Task Manager.

The error occurs at init.py line 857:
chat_handler_instance = handler_cls(clip_model_path=mmproj_path, verbose=False)

This triggers a C-level crash inside llama_cpp's clip_model_init(), which bypasses Python exception handling entirely, causing the whole process to hang.

Error message

ValueError: Failed to load clip model: mmproj-Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-BF16.gguf

Root cause

llama_cpp's clip_model_init() does not support the Qwen3.5 mmproj format. The crash happens at C++ level, so Python's try/except cannot catch it, resulting in a full process hang rather than a graceful error.

Suggestion

• Wrap the handler_cls() call in a try/except to prevent process crash (at minimum, catch BaseException or use subprocess isolation)
• Document the minimum llama-cpp-python version required for Qwen3.5 mmproj support
• Consider adding a version check before attempting to load the mmproj