Uploading LoRA adapter doesn't work

[p-e-w]

See https://huggingface.co/p-e-w/gpt-oss-20b-heretic-v3-LoRA/tree/main for an example. The adapter_model.safetensors file that is supposed to contain the LoRA is empty. Strangely enough, exporting the adapter to a local folder works, producing a 700 kB safetensors file.

[p-e-w]

Actually, it fails for all models. This sucks.

[kabachuha]

@p-e-w What do you think of saving the lora adapter to a temp dir and then upload it through hf hub upload folder?

https://huggingface.co/docs/huggingface_hub/v0.13.2/guides/upload#upload-a-folder

[p-e-w]

Too brittle, and not what the user expects when they choose "upload".

This has to work in PEFT/Transformers. We're not fixing upstream bugs downstream.

[kabachuha]

Idk add a disclaimer "local only" when saving a lora, just like the RAM disclaimer pops up

The lora file weights a few megabytes, it won't clutter the computer

You really are overcomplicating things

[kabachuha]

This is helpful for like 99%+ users who don't have 256+ RAM

[p-e-w]

I continue to believe that the correct way forward is to fix the problem, not work around it. That is, model.push_to_hub should do what it claims to do.

I imagine that the root cause of this bug is rather simple, and probably easy to resolve. It's just that this feature is completely useless to me personally, and I have more important things (for me) to work on at the moment.

[p-e-w]

The recommended workaround for people who need this feature is to revert 025ab3a. I will not merge a fix or workaround into master without understanding the cause of the problem.

[anrp]

I wonder what's broken? Maybe try again, was it a transient issue?
I tried tag v1.2.0 + revert 025ab3a, command line uv run heretic --model Qwen/Qwen3-0.6B --batch-size 64 --n-trials 10 --n-startup-trials 5 and uploaded lora adapter to https://huggingface.co/anrp/testhereticuploadlora and it worked fine, safetensors file is 1MB-ish?

Edit: to be clear, uploaded the lora from the menu options available, not saved & then uploaded.

[p-e-w]

@anrp

Thank you for bringing this to my attention, perhaps this has been fixed in Transformers then!

I will verify this soon, and if it indeed works reliably, will re-enable this option for the upcoming 1.3 release of Heretic.