Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
People complaining that they can't use the LoRA with VLLM because
load_lora
method is not available. This is because when loading a LoRA model,get_peft_model
goes into theUnsloth: Already have LoRA adapters! We shall skip this step.
patching.This PR is simply putting the patching inside that stage as well. We can't just move the patching to the beginning on the function or else when doing inference while training (like training GRPO), it'll do the inference only on the base model
This PR still has a flaw that the inference of vLLM has to be run once first before it's able to do inference on the loaded lora. Which maybe related of this part in the unsloth-zoo?
https://github.com/unslothai/unsloth-zoo/blob/a9857088bdaf412bef36800d837a3a37657555c8/unsloth_zoo/vllm_utils.py#L1206-L1212
Related issue -> #1670 (comment)