Apply quantization during DPO QLoRA #115
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
lewtun
commented
Feb 2, 2024
| @@ -1,12 +1,12 @@ | |||
| # Model arguments | |||
| model_name_or_path: alignment-handbook/zephyr-7b-sft-qlora | |||
| torch_dtype: float16 | |||
| torch_dtype: bfloat16 | |||
There was a problem hiding this comment.
I turns out that using bfloat16 makes a non-trivial difference to downstream perf! cc @nathan-az :)
|
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. |
lewtun
commented
Feb 2, 2024
| model_kwargs = dict( | ||
| revision=model_args.base_model_revision, | ||
| trust_remote_code=model_args.trust_remote_code, | ||
| use_flash_attention_2=model_args.use_flash_attention_2, | ||
| torch_dtype=torch_dtype, | ||
| use_cache=False if training_args.gradient_checkpointing else True, | ||
| device_map=get_kbit_device_map() if quantization_config is not None else None, | ||
| quantization_config=quantization_config, |
There was a problem hiding this comment.
Note that this approach of quantizing and then merging in DPOTrainer is what Tim Dettmers suggests: https://twitter.com/Tim_Dettmers/status/1694654191325573456
lewtun
commented
Feb 5, 2024
|
|
||
| # LoRA arguments | ||
| use_peft: true | ||
| load_in_4bit: true | ||
| lora_r: 16 | ||
| lora_alpha: 16 | ||
| lora_r: 128 |
There was a problem hiding this comment.
Tuning these hparams was necessary to get close to zephyr-7b-beta perf on MT-Bench
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This PR fixes a bug where we weren't quantising the base model with QLoRA during DPO and thus were actually doing LoRA instead.
Now we first quantise the base model in 4bit and load the SFT adapter (which later gets merged within the
DPOTrainer). Although this isn't as memory efficient as loading two adapters in a single base model (example), it does provide the flexibility to customise the QLoRA config.I find that with these settings MT-Bench yields a score of 7.212, which is ~0.1 lower than
zephyr-7b-betaand could likely be improved with a bit more tuning of hparams.