Compile bug: ValueError: Can not map tensor 'lm_head.biases' when converting Qwen3-8B (MLX fused LoRA) model

[tommyjakobsen]

Git commit

git reset --hard origin/master

Operating systems

Mac

GGML backends

CUDA

Problem description & steps to reproduce

I am attempting to convert a Qwen3-8B model that was fine-tuned with LoRA using MLX and then fused, to GGUF format using llama.cpp's convert_hf_to_gguf.py script.

python -m mlx_lm.convert --hf-path Qwen/Qwen3-8B --mlx-path ./qwen3_8b_mlx_quantized -q

uv run mlx_lm.lora --train --model ./qwen3_8b_mlx_quantized --data ./data --batch-size 2 --iter 2000 (mention that adapters.safetensors was produced in ./adapters directory)

uv run python -m mlx_lm.fuse --model ./qwen3_8b_mlx_quantized --adapter-path ./adapters --save-path ./merged_qwen3_8b_mlx_fused --hf-path Qwen/Qwen3-8B

uv run python ../llama.cpp/convert_hf_to_gguf.py ./merged_qwen3_8b_mlx_fused --outfile qwen3_euaiact_8b_fused.gguf --outtype f16

The convert_hf_to_gguf.py script fails with a ValueError: Can not map tensor 'lm_head.biases'."

"I have verified that the lm_head.biases and lm_head.scales tensors exist in the model's model.safetensors.index.json (even in the fused model), which matches the original Qwen3 Hugging Face model structure."

Paste the output of cat merged_qwen3_8b_mlx_fused/model.safetensors.index.json | grep "lm_head":
JSON

"lm_head.biases": "model.safetensors",
"lm_head.scales": "model.safetensors",
"lm_head.weight": "model.safetensors",

config.json

uv run python ../llama.cpp/convert_hf_to_gguf.py ./merged_qwen3_8b_mlx_fused --outfile

First Bad Commit

No response

Compile command

cmake worked.

Relevant log output

qwen3_euaiact_8b_fused.gguf --outtype f16
INFO:hf-to-gguf:Loading model: merged_qwen3_8b_mlx_fused
INFO:hf-to-gguf:Model architecture: Qwen3ForCausalLM
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...
INFO:hf-to-gguf:gguf: loading model weight map from 'model.safetensors.index.json'
INFO:hf-to-gguf:gguf: loading model part 'model.safetensors'
Traceback (most recent call last):
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 6718, in <module>
    main()
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 6712, in main
    model_instance.write()
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 409, in write
    self.prepare_tensors()
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 277, in prepare_tensors
    for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 2743, in modify_tensors
    yield from super().modify_tensors(data_torch, name, bid)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 245, in modify_tensors
    return [(self.map_tensor_name(name), data_torch)]
             ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/tommy.jakobsen/Documents/_DEMOS/_Ai_agents/finetune/euaiact/../llama.cpp/convert_hf_to_gguf.py", line 236, in map_tensor_name
    raise ValueError(f"Can not map tensor {name!r}")
ValueError: Can not map tensor 'lm_head.biases

[CISC]

The scales and biases tensors are unknown entities, so the conversion script does not know what to do with them, if you do feel free to make a PR. :)

[nthsense]

Hit the exact same issue... fine-tuned Qwen 3 using MLX, fused the model, and am trying to convert to GGUF. For me, the issue is that when I fused the model, it created a quantized version of the model. The scales and biases tensors are there (as I understand it) to help reconstruct the original range of the quantized tensors. So the solution for me was to run mlx_lm fuse with --de-quantize and then convert to GGUF.