nvfp4 gguf of Qwen 3.5 122B

[xkmire]

Not sure anyone wants more patches to the convert_hf_to_gguf.py. The file is big as it is.
However, I wanted a NVFP4 variant of the latest Qwen3.5-122B (huggingface: txn545/Qwen3.5-122B-A10B-NVFP4), that was produced with NVIDIA Model Optimizer. However, the organization of tensors was slightly different than the existing script.

Anyway, if anyone needs a NVFP of this model, below is a patch.

It will be really nice when GPU support comes for NVFP4 in llama.cpp. For now, it runs mostly on the CPU (but some parts look to be running on the GPU, which I assume is the non NVFP4 parts of calculations).

[xkmire]

diff.txt