Add support for QLoRA/ QAdapter training via bitsandbytes #663
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
4 tasks
4 tasks
calpt
changed the title
hSterz
approved these changes
Apr 22, 2024
| # result shape: <batch_size> x <seq_len> x <head_dim> | ||
| layer_output = F.linear(input_states, weight, bias=self.bias) | ||
| else: | ||
| layer_output = super().forward(input_states) |
There was a problem hiding this comment.
Which forward method is called here since this does not inherit from nn.Linear anymore?
There was a problem hiding this comment.
the subclasses of this (LoRALinearTorch, LoRALinear4bit, LoRALinear8bitLt), inherit from different types of linear layers
calpt
added a commit
that referenced
this pull request
Apr 22, 2024
…678) Adapters currently does not work correctly with passing `device_map="auto"` in a model's `from_pretrained()`. Device auto-mapping is handled by HF accelerate, which wraps the original module forward method. This PR fixes compatibility of Adapters' post-hoc model wrapping with Accelerate's device auto-mapping via wrapping the forward pass. Fixing this is required for enabling quantized training of adapters (bottleneck & prefix-tuning) in #663.
calpt
changed the title
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Outdated suggestions cannot be applied.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 adds support for wrapping bitsandbytes'
Linear4bitandLinear8bitLtquantization layers with our LoRA implementation, enabling training LoRA adapters on quantized models in QLoRA style.Implementation is loosely similar to HF peft's approach, which can be found here: https://github.com/huggingface/peft/blob/v0.10.0/src/peft/tuners/lora/bnb.py.
Demo
I've added a new notebook here: https://github.com/calpt/adapter-transformers/blob/dev/qlora/notebooks/QLoRA_Llama_Finetuning.ipynb.
The notebook showcases this feature by finetuning a 4bit-quantized Llama 2 7B on an instruction tuning dataset (similar to Guanaco in the QLoRA paper).
Tested that it runs without errors in the provided notebook, other setups are not extensively tested yet.
Pre-trained checkpoints
Adapters trained with the notebook code can be found here:
Llama-2 7B: https://huggingface.co/AdapterHub/llama2-7b-qlora-openassistant
Llama-2 13B: https://huggingface.co/AdapterHub/llama2-13b-qlora-openassistant
Current limitations