convert_lora_to_gguf: support Qwen3.5/3.8 multimodal LoRA

[the-pgh-cid]

Overview

Fix convert_lora_to_gguf.py so a LoRA trained on a Qwen3.5/3.8 multimodal base (architecture Qwen3_5ForConditionalGeneration, e.g. Qwen/Qwen3.8-27B) converts to GGUF instead of failing.

The failure: ValueError: Can not map tensor 'model.language_model.layers.11.self_attn.k_proj.weight'

Root cause: multimodal conditional-generation bases nest the language stack under model.language_model.layers.N.... get_base_tensor_name() strips only base_model.model., so the mapped name still has the language_model. segment, which the base tensor map does not recognize. The base converter (convert_hf_to_gguf.py) already strips language_model. when building the base GGUF; the LoRA converter never applies the same strip.

The change is one line in get_base_tensor_name:

base_name = base_name.replace("model.language_model.", "model.")

It applies only to the exact prefix the base converter already strips. Dense (non-multimodal) models have no language_model. in their names, so they are unaffected.

Additional information

Verified on Qwen/Qwen3.8-27B: a PEFT LoRA (r=16, attention projections only) with 128 tensors converts cleanly after the fix. Before the fix the same command fails with the error above. The adapter (F16 GGUF, 21MB, 128 tensors) loads and serves against the base.

Reproduction:

python convert_lora_to_gguf.py <adapter_dir> \
  --base /path/to/Qwen3.8-27B-HF --outfile out.gguf

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: YES. I diagnosed the root cause with assistance from an AI coding agent; the fix is a one-line tensor-name strip that I reviewed, tested, and own. Per AGENTS.md, I am responsible for this change and can discuss it independently.

[ggml-gh-bot]

Hi @the-pgh-cid, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[CISC]

Are you using the latest script? Just as in the base converter lora conversion goes through the same tensor filtering and language_model gets filtered here:

llama.cpp/conversion/base.py

Lines 583 to 584 in 6efc773

	if "language_model." in name:
	name = name.replace("language_model.", "")

[the-pgh-cid]

It is very likely that I am using an older script. Now, with pie on my face, I will head back and check and see if my brain is also plugged in, so to speak lol. thank you. I was absolutely using 9138 from a June 16 pull. My current fork supercedes my predated cuda build, and I never thought to look there first.