convert : support mixed-precision ModelOpt models with per-tensor NVFP4/FP8 quantization

[richarddd]

Adds support for converting mixed-precision ModelOpt models (e.g. nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4) that use per-tensor quant_algo with both NVFP4 and FP8 layers, instead of a single global quant_algo: "NVFP4". NVFP4 tensors (2D scales) are repacked natively while FP8 tensors (1D scales) are dequantized to float.

Fixes: #20504

[vbooka1]

Either llama.cpp or convert_hf_to_gguf.py is broken, model https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 does not work with builds 8304 (first supporting Nemotron 3) and 8334 (latest): I'm getting error llama_model_load: error loading model: done_getting_tensors: wrong number of tensors; expected 844, got 843.

Please check.

convert.txt

launch.txt

[richarddd]

Either llama.cpp or convert_hf_to_gguf.py is broken, model https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 does not work with builds 8304 (first supporting Nemotron 3) and 8334 (latest): I'm getting error llama_model_load: error loading model: done_getting_tensors: wrong number of tensors; expected 844, got 843.

Please check.

convert.txt

launch.txt

Yeah something is off. I didn’t properly smoke test due to lack of memory.

[CISC]

@vbooka1 @richarddd Fixed by #20506