convert: fix Nemotron-H LoRA GGUF conversion

[frozenblade1224]

Overview

LoRA adapters cannot be converted to GGUF for Nemotron-H models. convert_lora_to_gguf.py fails with two one independent, sequential defects.

Defect 1 - hparams resolution (Resolved in #27101 )

Defect 2 - After working around Defect 1, the first adapter tensor is rejected. PEFT names adapter tensors from the runtime module tree (model.layers.*), while gguf-py/gguf/tensor_mapping.py registers only the on-disk checkpoint form (backbone.layers.*) for Nemotron-H. The two naming domains never meet.

  File "C:\---------\llama.cpp\conversion\base.py", line 617, in map_tensor_name
    raise ValueError(f"Can not map tensor {name!r}")
ValueError: Can not map tensor 'model.layers.12.mixer.k_proj.weight'

Fix 2 - We rename model.layers.* to backbone.layers.* for tensors containing .mixer. before delegating to the parent filter.

Additional information

Defect 2 cannot be reached without Defect 1's resolution.

The test command used for all cases is -

py .\convert_lora_to_gguf.py    --base ..\NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16   --outtype f16   --outfile ..\adapter.gguf   "C:\---------\adapters\lora_r8_re_zeros"

The used LORA adaptor is available here. After applying the 2 fixes identified above, we can get the following successful result -

INFO:lora-to-gguf:Model successfully exported to ..\adapter.gguf

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, for code review and analysis.

[CISC]

Point 1 should be fixed by #27101, point 2 is a bit mysterious, why do the naming differ?

[frozenblade1224]

Indeed, #27101 settles point 1.

As for point 2, the model.\* in the tensor naming arises from NemotronHForCausalLM.__init__ in this repo. PEFT finds LoRA targets by looking through the live PyTorch module hierarchy, which consequently emits model.layers.12.mixer.k_proj.weight.

The found name disagrees with the backbone.layers.12.mixer.k_proj.weight mapped in gguf-py/gguf/tensor_mapping.py (and the actual namespace in the base checkpoint). This mismatch causes the error.

[frozenblade1224]

@CISC I have removed the changes I made for Defect 1. The push now solely handles Defect 2.