1. unsloth updated in the image two days ago (December 6)
2. using vast.ai with my own image
3. One 3090Ti in this instance, but does not depend on GPU; I use a single GPU
4. No notebook, pure ssh
5. unsloth 2025.11.6, transformers 4.57.1, trl 0.24.0, torch 2.9.1, peft 0.18.0
6. regular Trainer

I want to use CorDA: https://github.com/huggingface/peft/tree/main/examples/corda_finetuning

In CorDA, you initialize LoraConfig with (minimal example from the link above):

corda_config = CordaConfig(
    corda_method="kpm",
)
lora_config = LoraConfig(
    init_lora_weights="corda",
    corda_config=corda_config,
)

With Unsloth, FastLanguageModel.get_peft_model() encapsulates LoraConfig and does not allow the passing of init_lora_weights="corda" (one can still pass corda_config via kwargs). This makes CorDA unusable without workarounds, such as applying eigens manually after get_peft_model()