mlx-lm v0.28.4 failure

#1
by Epistates - opened

This model looks really interesting for coding! I tried to use:
mlx_lm.convert --hf-path DavidAU/Qwen3-48B-A4B-Deadpan-Savant-12x-Closed-Open-Source-Distill --mlx-path /mypath/Qwen3-48B-A4B-Deadpan-Savant-12x-Closed-Open-Source-Distill

Normally Qwen3MoeForCausalLM converts without issue but this model fails with:

mlx/nn/layers/base.py", line 191, in load_weights
    raise ValueError(f"Missing {num_missing} parameters: \n{missing}.")
ValueError: Missing 1 parameters:
lm_head.weight.

https://github.com/ml-explore/mlx/blob/main/python/mlx/nn/layers/base.py#L188-L191

Hey ;

The head/embed are tied on this MOE; which is the cause of this error.
I will attempt to "untie" them next version ; however all the expert models must also be "untied" too.

I don't know of a way to bypass this MLX issue.
If you find a way ; please share as this will help with other models too.

thanks'
David

Epistates changed discussion status to closed

Sign up or log in to comment