Custom Model support

[huangtingwei9988]

hi～，@YangWang92 ，for custom model, how to generate the hessian and inv_hessian weights?

[YangWang92]

Hi @huangtingwei9988, thank you for your interest in our method.

The Hessian matrix can be collected through quip-sharp (here is my fork, which includes some minor fixes): https://github.com/YangWang92/quip-sharp/blob/wy/hessian/note.md and #79 .

The inverse Hessian can be obtained from https://gist.github.com/YangWang92/ec98a86c3a33c573b601cf4348d0a0e7.

The current tutorial is a bit simple, but I will release a complete tutorial on obtaining the Hessian soon. Please stay tuned.

[huangtingwei9988]

Hi @huangtingwei9988, thank you for your interest in our method.

The Hessian matrix can be collected through quip-sharp (here is my fork, which includes some minor fixes): https://github.com/YangWang92/quip-sharp/blob/wy/hessian/note.md and #79 .

The inverse Hessian can be obtained from https://gist.github.com/YangWang92/ec98a86c3a33c573b601cf4348d0a0e7.

The current tutorial is a bit simple, but I will release a complete tutorial on obtaining the Hessian soon. Please stay tuned.

Thank you very much. I also want to know whether the qwen2-moe model will be supported in the future?

[YangWang92]

Hi @huangtingwei9988, thank you for your interest in our method.
The Hessian matrix can be collected through quip-sharp (here is my fork, which includes some minor fixes): https://github.com/YangWang92/quip-sharp/blob/wy/hessian/note.md and #79 .
The inverse Hessian can be obtained from https://gist.github.com/YangWang92/ec98a86c3a33c573b601cf4348d0a0e7.
The current tutorial is a bit simple, but I will release a complete tutorial on obtaining the Hessian soon. Please stay tuned.

Thank you very much. I also want to know whether the qwen2-moe model will be supported in the future?

I previously have a quick try on Qwen2.5 A52 and encountered some minor issues. I haven't had the time to fix them yet. Please give me 1-2 weeks, and I will address the issues in MoE models.