Can't convert mistral 3 large

[bartowski1182]

I get the following error:

ValueError: Can not map tensor 'layers.41.experts.66.w1.weight'

Using the --mistral-format arg (without it it fails even earlier)

All other sizes converted without issue and without the argument (presumably because they include a config.json)

[KyleBruene]

I would assume we'll have to wait until it gets into transformers.

"We sadly didn't have enough time to add Mistral Large 3 to transformers, but we would be very happy for a community contribution by opening a PR to huggingface/transformers."

https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512

[bartowski1182]

I would think that the --mistral-format would work, since I think that's the goal of it, but may be something missing

[DocShotgun]

The arch itself is essentially a DeepSeekV3 with a vision tower, so in theory it shouldn't be too difficult to make work in llama.cpp?

vllm-project/vllm#29757

[bartowski1182]

#17730 is working so far, doesn't perfectly support the new model allegedly but does get conversion unlocked!