Fine tuning with Lora does not support this model

#4
by fromjon - opened
MLX Community org
edited Aug 13

I attempt to run the lora.py trainer with this model but it fails to start

...mlx_lm/tuner/utils.py", line 129, in linear_to_lora_layers
    raise ValueError(f"Lora does not support {model.model_type}")
ValueError: Lora does not support deepseek_v2

Can mlx_lm fine-tune models without LoRA?
If so, how?

MLX Community org

The DeepSeek V2 is using MLA attention. I haven't tried LoRA yet, so it hasn't been added to MLX_LM's LoRA support. It might be better to open an issue on the mlx-example repo to see how to add LoRA support for it.

MLX Community org

@fromjon This has been fixed by https://github.com/ml-explore/mlx-examples/pull/932. Let me know if it still doesnt work for you

mzbac changed discussion status to closed

Sign up or log in to comment