Fine-tuning without LoRA

[bensonbs]

1. Is this request related to a challenge you're experiencing? Tell us your story.

I would like to understand the process and feasibility of fine-tuning an entire model instead of using LoRA (Low-Rank Adaptation). While LoRA is great for parameter-efficient fine-tuning, in my case, I am exploring scenarios where I need to fine-tune the entire model to achieve better control over its performance.

---

2. What is your suggested solution?

I propose modifying the training code as follows:

model = BaseTransformer.from_pretrained(
    path="path_to_your_pretrained_model",
    load_weights=True,
    lora_config=None,  # <-- Pass None here
)

Then, execute the training command:

python fish_speech/train.py --config-name text2semantic_finetune \\
 project=$project \\

If training without LoRA, is it necessary to convert the LoRA weights back to regular weights?

What steps should I follow before inference?

---

3. Additional context or comments

No response.

---

4. Can you help us with this feature?

• I am interested in contributing to this feature.

[PoTaTo-Mika]

pr welcome for full params finetune.

[abhisirka2001]

Could you do full param finetuning of fishspeech?