Support for GLM?

[silk isle]

I'm absolutely loving GLM4-0414 right now and want to finetune it for and do GPRO using unsloth, but realized that GLM has a different chat template (unless i'm overthinking it and i can just use like shareGPT?). Do you guys support GLM in the first place or is it just a matter of making a new template?

Thanks for all the work you guys do by the way, you are all absolute G's

[radiant whale]

GLM4 is great model! I am now using it exclusively, 9B and 32B versions.

[night crag]

We support GLM yes and you'll need a chat template

We also uploaded dynamic 2.0 GGUFs here: https://huggingface.co/unsloth/GLM-4-32B-0414-GGUF

[silk isle]

Oooh any chance you guys could upload your magic quants for the base model? like the unsloth-bnb-4bit safetensors?

I managed to get a fine tune going with a good loss curve but got this when running the fibonnaci sequence code block

The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's attention_mask to obtain reliable results.

[silk isle]

bumping, got this warning.

so far it seems like im getting it to work!

[silk isle]

@night crag @delicate beacon Whenever you guys get the time 🙂 No need to fix now but just something to look out for

[night crag]

Is it fixed in the latest Unsloth update? 🫣

Apologies btw just saw this

[silk isle]

No worries! I’m gonna try another run this weekend 🙂

Still struggling with getting a working chat template for Z1 though

there’s some slight differences with the GLM4 template with some logic related to the think tags