README.md · sayakpaul/flux-lora-resizing at c392593afed3289482c495a0e02df17f5ecfbe92

metadata

language:
  - en
library_name: diffusers
license: other
license_name: flux-1-dev-non-commercial-license
license_link: LICENSE.md

LoRA is the de-facto technique for quickly adapting a pre-trained large model on custom use cases. Typically, LoRA matrices are low-rank in nature. Now, the word “low” can vary depending on the context, but usually, for a large diffusion model like Flux, a rank of 128 can be considered high. This is because users may often need to keep multiple LoRAs unfused in memory to be able to quickly switch between them. So, the higher the rank, the higher the memory on top of the volume of the base model.

So, what if we could take an existing LoRA checkpoint with a high rank and reduce its rank even further to:

Reduce the memory requirements
Enable use cases like torch.compile() (which require all the LoRAs to be of the same rank to avoid re-compilation)

This project explores two options to reduce the original LoRA checkpoint into an even smaller one:

Random projections
SVD

Random projections

Basic idea:

Generate a random projection matrix: R = torch.randn(new_rank, original_rank, dtype=torch.float32) / torch.sqrt(torch.tensor(new_rank, dtype=torch.float32)).
Then compute the new LoRA up and down matrices:
```
# We keep R in torch.float32 for numerical stability.
lora_A_new = (R @ lora_A.to(R.dtype)).to(lora_A.dtype)
lora_B_new = (lora_B.to(R.dtype) @ R.T).to(lora_B.dtype)
```
If lora_A and lora_B had shapes of (42, 3072) and (3072, 42) respectively, lora_A_new and lora_B_new will have (4, 3072) and (3072, 4), respectively.

Results

Tried on this LoRA: https://huggingface.co/glif/how2draw. Unless explicitly specified, a rank of 4 was used for all experiments. Here’s a side-by-side comparison of the original and the reduced LoRAs (on the same seed).

Inference code

from diffusers import DiffusionPipeline 
import torch 

pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
# Change accordingly.
lora_id = "How2Draw-V2_000002800_svd.safetensors"
pipe.load_lora_weights(lora_id)

prompts = [
    "Yorkshire Terrier with smile, How2Draw",
    "a dolphin, How2Draw",
    "an owl, How3Draw",
    "A silhouette of a girl performing a ballet pose, with elegant lines to suggest grace and movement. The background can include simple outlines of ballet shoes and a music note. The image should convey elegance and poise in a minimalistic style, How2Draw"
]
images = pipe(
    prompts, num_inference_steps=50, max_sequence_length=512, guidance_scale=3.5, generator=torch.manual_seed(0)
).images

	Yorkshire Terrier with smile, How2Draw
	a dolphin, How2Draw
	an owl, How3Draw
	A silhouette of a girl performing a ballet pose, with elegant lines to suggest grace and movement. The background can include simple outlines of ballet shoes and a music note. The image should convey elegance and poise in a minimalistic style, How2Draw

Code: low_rank_lora.py

Notes

One should experiment with the new_rank parameter to obtain the desired trade-off between performance and memory. With a new_rank of 4, we reduce the size of the LoRA from 451MB to 42MB.
There is a use_sparse option in the script above for using sparse random projection matrices.

SVD

Results

Randomized SVD

Full SVD can be time-consuming. Truncated SVD is useful very large sparse matrices. We can use randomized SVD for none-to-negligible loss in quality but significantly faster speed.

Results

Code: svd_low_rank_lora.py

Tune the knobs in SVD

new_rank as always
niter when using randomized SVD

Reduced checkpoints

Randomized SVD: How2Draw-V2_000002800_rand_svd.safetensors
Full SVD: How2Draw-V2_000002800_svd.safetensors
Random projections: How2Draw-V2_000002800_reduced.safetensors