Why loading a lora weights so low?

[zengjie617789]

I used diffusers to load lora weights but it much slow to finish.
diffusers version: 0.29.2

I test another version of diffusers 0.23.0 without peft installation, and the time is decent.

t1 = time.time()
pipe.load_lora_weights("/data/**/lora_weights/lcm-lora-sdxl/", weight_name="pytorch_lora_weights.safetensors")
print(f"load lcm lora weights cost: {time.time()- t1}")

And If I use low version of diffusers, much of code need to be modified which cost much work.
Anyone who can help me will be appreciate.

[tolgacangoz]

Hmm, you seem right. I observed similar 3-5 times differences in speed between 0.27<=diffusers<=0.29.1 w peft and 0.23<=diffusers<=0.26 w/o peft. At version diffusers==0.27, peft was set to compulsory. peft really seems to could add overhead 🤔.

[a-r-r-o-w]

cc @sayakpaul

[sayakpaul]

Cc: @BenjaminBossan

[BenjaminBossan]

Could you please provide a full code example so that I can try to reproduce these timings? Also, what version of PEFT do you have installed?

[zengjie617789]

@BenjaminBossan The code are plain code to load model and run inference, and I load LCM lora weight as a test as the same as using other loras. Code snippets below like:

pipe = StableDiffusionXLPipeline.from_pretrained(model_path,  torch_dtype=torch.float16, variant="fp16", use_safetensors=True).to(device)
t1 = time.time()
pipe.load_lora_weights("/data/***/models/lora_weights/lcm-lora-sdxl/", weight_name="pytorch_lora_weights.safetensors")
print(f"load lcm lora weights cost: {time.time()- t1}")

diffusers==0.25.1 peft==0.12.0
the gap of time are because of the installation of peft or not.

[BenjaminBossan]

Thanks for providing more context. Since you use a private adapter, I changed the code a little bit to use one from the Hub:

import time

import torch
from diffusers import StableDiffusionXLPipeline

try:
    import peft
    print("peft is installed")
    print(peft.__version__)
    print(peft.__path__)
except ImportError:
    print("peft is not installed")

device = 0
model_path = "stabilityai/stable-diffusion-xl-base-1.0"

pipe = StableDiffusionXLPipeline.from_pretrained(
    model_path, torch_dtype=torch.float16, variant="fp16", use_safetensors=True,
).to(device)
t1 = time.time()
pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl", weight_name="pytorch_lora_weights.safetensors")
print(f"load lcm lora weights time: {time.time()- t1:.2f}s")

With this, I can confirm these numbers (averaged over 5 runs):

• diffusers v0.23.0, no PEFT: ~0.7s
• diffusers v0.23.0, PEFT v0.12.0: ~2.0s
• diffusers v0.29.0, PEFT v.012.0: ~2.0s

I investigated further why PEFT is slower and what I found is that with PEFT, when the LoRA layer is created (before the actual checkpoint is loaded), PEFT will instantiate actual LoRA weights. Without PEFT, empty weights are created (meta device).

Next I tried to turn off using empty weights to check if that really makes the difference. However, when passing low_cpu_mem_usage=False to load_lora_weights, which from my understanding controls this behavior, I found that the value is not correctly passed down, so that here, it is set to True again. Not sure if this is intentional or not. To get it to work, I had to hard-code low_cpu_mem_usage=False in this line.

Long story short, after forcing this, I get very similar load times of ~1.7s. So in a sense this is a regression, as using the PEFT backend in diffusers is not loading empty weights. I'm not quite sure if this is an oversight, a bug, or intended for some reason. @sayakpaul do you think this could be added for PEFT LoRA loading?

[sayakpaul]

We had brought it up in the past a couple of times but as you can see the difference is not really that significant so, I am not sure about the value add here given the number of lines of code to be added to PEFT.

[sayakpaul]

I also don’t understand what was incorrectly passed down for low_cpu_mem_usage in the function you mentioned. Do you mean the value was overridden when you passed low_cpu_mem_usage=False?

[BenjaminBossan]

We had brought it up in the past a couple of times but as you can see the difference is not really that significant so, I am not sure about the value add here given the number of lines of code to be added to PEFT.

Well, it's up to you and the other diffusers maintainers if you think this is worth fixing or not. On the PEFT side, I think it could be a feature to allow passing some argument to inject_adapter_in_model to tell PEFT to load the weights only on meta device, then this change would be easy to make on the diffusers side. I'll put it on the backlog and can notify you once we the feature is added.

I also don’t understand what was incorrectly passed down for low_cpu_mem_usage in the function you mentioned. Do you mean the value was overridden when you passed low_cpu_mem_usage=False?

The value is not passed on in this line:

https://github.com/huggingface/diffusers/blob/v0.23.0/src/diffusers/loaders.py#L3234

It is set correctly in the kwargs but is not passed to load_lora_into_unet (not sure if intentional or not). As this is v0.23.0, there is probably not any TODO from this finding, even if it's an error.

[sayakpaul]

Ah you are right. But we have moved from this already. So, I think we can ignore it for now.

But on a second thought, I think faster loading support would indeed be nice especially for larger adapters. Please let us know when this has been shipper in PEFT and we can do the necessary changes in diffusers.

[BenjaminBossan]

I did some quick and dirty experiments and created a draft PR based on those. There is indeed a speed up for Llama3 8B in this test, although loading the base model is still the overall bottleneck (even with hot cache and M.2 SSD).

Overall, I thus think the impact of this feature will not be huge, especially since this a one time cost, but implementing it is not trivial. Therefore, I'll leave this in a draft PR stage for the time being while working on higher priority items.

[sayakpaul]

This is what I had expected too. Thanks for quickly checking!

[zengjie617789]

Thank you for you all. I test the version as refered, the version are as below:

but the time cost much long,

If I test it without peft:

The nvidia card is A100.
Obviously, the difference between with peft or not is huge. And I wonder how to get the result of 0.7s vs 2s.

[BenjaminBossan]

I test the version as refered, the version are as below:

Just to be clear, the test that I mentioned is in a draft PR, so it is not available on PEFT yet. And even if you installed PEFT from my branch, it would still not work because diffusers would need to make some changes too to opt into that feature. I will get back to that PR when I have a bit of extra time.

And I wonder how to get the result of 0.7s vs 2s.

These were based on another adapter ("latent-consistency/lcm-lora-sdxl") since you seem to use a private one that I cannot load. Also, a lot depends on the hardware. When you run the snippet I posted above, what times do you get? Please run a couple of times to warm up the cache and lower the variance.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[BenjaminBossan]

not stale, huggingface/peft#1961 is being worked on

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[sayakpaul]

With #9510 and the latest versions of peft and transformers (installed from main as of now), this should be significantly improved. So, I am going to close this. Feel free to reopen.