Feat: add support for Conv2D DoRA

[sayakpaul]

TODO

• Incorporate _apply_dora() in the forward() of Conv2D
• merge()
• unmerge()
• Tests
• Docs

[HuggingFaceDocBuilderDev]

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.

[sayakpaul]

@BenjaminBossan would love to get your thoughts on the changes I have introduced so far here.

[BenjaminBossan]

Thanks a lot Sayak, this already looks pretty good. Probably the main thing to get right is the dimensions across which to normalize, which does look correct to me. Maybe there is also some room to speed up the application of the norm, not sure.

As a next step, let's add a test case for Conv2d + DoRA. This should be as easy as adding an example or two here:

peft/tests/test_custom_models.py

Lines 58 to 70 in 7e84dec

	("Vanilla MLP 7 LoRA with DoRA", "MLP", LoraConfig, {"target_modules": ["lin0"], "use_dora": True}),
	("Vanilla MLP 8 LoRA with DoRA", "MLP", LoraConfig, {"target_modules": ["lin0", "lin1"], "use_dora": True}),
	(
	"Vanilla MLP 9 LoRA with DoRA",
	"MLP",
	LoraConfig,
	{"target_modules": "lin1", "use_dora": True, "lora_alpha": 32},
	),
	("Embedding + transformers Conv1D 1 LoRA", "EmbConv1D", LoraConfig, {"target_modules": ["conv1d"]}),
	("Embedding + transformers Conv1D 2 LoRA", "EmbConv1D", LoraConfig, {"target_modules": ["emb"]}),
	("Embedding + transformers Conv1D 3 LoRA", "EmbConv1D", LoraConfig, {"target_modules": ["emb", "conv1d"]}),
	("Conv2d 1 LoRA", "Conv2d", LoraConfig, {"target_modules": ["conv2d"]}),
	("Conv2d 2 LoRA", "Conv2d", LoraConfig, {"target_modules": ["conv2d", "lin0"]}),

Let's do this and check if the tests pass, then plan the next steps.

[sayakpaul]

@BenjaminBossan I think this is ready for a review now.

[sayakpaul]

Cc: @nbasyl if you want to give this a look as well.

[BenjaminBossan]

This looks pretty good, thanks a lot for enabling DoRA with Conv2d. Curious to see if we can get nice improvements for diffusion models.

Implementation-wise, I have only a few minor comments. On top of those, could you please document the new changes here:

• The docstring of use_dora inside of tuners/lora/config.py still mentions that only linear layers are supported.
• Same for the help of use_dora in the same file.
• docs/source/developer_guide/lora.md also needs updating with regard to the supported layers.

Before merging, it would be nice to have another review.

Edit: Please also merge with recent main, as there could be merge conflicts with #1518 otherwise.

[sayakpaul]

@BenjaminBossan resolved your comments. Thank you!

[BenjaminBossan]

Thanks a lot, Sayak, the PR looks great. Let's see if Shih-yang also has time to take a look. Otherwise, I'd be glad to have a second pair of eyes by @younesbelkada or @pacman100.

[nbasyl]

The PR looks great, I have skimmed through the code and did not notice any problems. @sayakpaul, Thanks for the effort!

[BenjaminBossan]

Thanks so much Sayak.