[Docs] Add unsloth optimizations in TRL's documentation
#1119
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
Cool work @younesbelkada and thanks! Looks good - I'll add a few changes in a bit! |
docs/source/sft_trainer.mdx
Outdated
|
|
||
| ### Accelerate fine-tuning using `unsloth` library | ||
|
|
||
| You can further accelerate QLoRA / LoRA and even full-finetuning using [`unsloth`](https://github.com/unslothai/unsloth) library that is compatible with `SFTTrainer`. Currently `unsloth` supports only Llama and Mistral architectures. |
There was a problem hiding this comment.
"You can further accelerate QLoRA / LoRA and even full-finetuning using" maybe to
You can further accelerate QLoRA / LoRA 2.2x faster and use 60% less memroy and even full-finetuning (1.1x speed boost) using unsloth library that is compatible with SFTTrainer. Currently unsloth supports only Llama (Yi, TinyLlama etc) and Mistral architectures.
docs/source/sft_trainer.mdx
Outdated
| @@ -410,6 +410,61 @@ We have tested NEFTune by training `mistralai/Mistral-7B-v0.1` on the [OpenAssis | |||
| </div> | |||
|
|
|||
| Note however, that the amount of performance gain is _dataset dependent_ and in particular, applying NEFTune on synthetic datasets like [UltraChat](https://huggingface.co/datasets/stingning/ultrachat) typically produces smaller gains. | |||
|
|
|||
| ### Accelerate fine-tuning using `unsloth` library | |||
There was a problem hiding this comment.
Accelerate fine-tuning using unsloth library
maybe to
Accelerate fine-tuning using unsloth
|
Actually instead of comments, I just editted the docs in place here: add-unsloth-docs...danielhanchen:trl:patch-1 |
Co-authored-by: Daniel Han <[email protected]>
|
Thanks a lot @danielhanchen ! I just merged your PR with your suggestions |
|
:) |
…e#1119) * add unsloth * Update sft_trainer.mdx (huggingface#1124) Co-authored-by: Daniel Han <[email protected]> --------- Co-authored-by: Daniel Han <[email protected]>
What does this PR do?
This PR adds details about unsloth library: https://github.com/unslothai/unsloth that can bring very interesting out of the box speedups to users. In the future, I'll make sure users that use unsloth + SFTTrainer have the tag
unslothwhen they push models on the hub with unslothcc @danielhanchen @lvwerra
@danielhanchen let me know if the content does look to you! Feel free to add any suggestion directly in the PR
Related: unslothai/unsloth#34