CompVis

/

stable-diffusion-v1-4

like 6.3k

We're working on exactly this! Pinging @pcuenq and @apolinario here as well

Please also check announcements on Twitter - we'll publish something about that soon!

Thanks!

In my case on Apple M1 with the code

# make sure you're logged in with `huggingface-cli login`
import os
import torch
from torch import autocast
from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler

# To swap out the noise scheduler, pass it to from_pretrained:
lms = LMSDiscreteScheduler(
    beta_start=0.00085, 
    beta_end=0.012, 
    beta_schedule="scaled_linear"
)

device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f'running on {device}')
pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-3", 
    scheduler=lms,
    torch_dtype=torch.float16,
    revision="fp16",
    use_auth_token=True,
    cache_dir=os.getenv("cache_dir", "./models")
).to(device)

prompt = "a photo of an astronaut riding a horse on mars"
with autocast(device):
    image = pipe(prompt)["sample"][0]  
    
image.save("astronaut_rides_horse.png")

I get the following error

Traceback (most recent call last):
  File "diffuser.py", line 27, in <module>
    image = pipe(prompt)["sample"][0]  
  ...
  File "/Documents/Projects/bloom/.venv/lib/python3.7/site-packages/torch/nn/functional.py", line 2503, in layer_norm
    return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

For device mps it doesn't work out of the box yet indeed, however if device = cpu it should work @loretoparisi . Can you try removing the with autocast(device)? Autocast doesn't work for CPU as of now

For device mps it doesn't work out of the box yet indeed, however if device = cpu it should work @loretoparisi . Can you try removing the with autocast(device)? Autocast doesn't work for CPU as of now

Thanks, I slightly modified the code like

prompt = "a photo of an astronaut riding a horse on mars"
samples = 2
steps = 45
scale = 7.5
if device=='cuda':
    with autocast(device):
        image = pipe(
            [prompt]*samples,
            num_inference_steps=steps,
            guidance_scale=scale,
            )["sample"][0]
else:
    image = pipe(prompt)["sample"][0]

but I'm still getting the same error:

Traceback (most recent call last):
  File "diffuser.py", line 39, in <module>
    image = pipe(prompt)["sample"][0]
  File "/Projects/bloom/.venv/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
...
  File "/Projects/bloom/.venv/lib/python3.7/site-packages/torch/nn/functional.py", line 2503, in layer_norm
    return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

@loretoparisi , oh this is probably because you are trying to load the fp16 version of the model - which also doesn't work on CPU 😅
Try this for pipe

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", #better model btw 
    scheduler=lms,
    use_auth_token=True,
    cache_dir=os.getenv("cache_dir", "./models")
).to(device)

Thank you it works without on Apple M1, removing autocast and fp16!

Here is the code for other people's convenience

# make sure you're logged in with `huggingface-cli login`
import os
import torch
from torch import autocast
from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler

# To swap out the noise scheduler, pass it to from_pretrained:
lms = LMSDiscreteScheduler(
    beta_start=0.00085, 
    beta_end=0.012, 
    beta_schedule="scaled_linear"
)

device = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f'running on {device}')

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", #better model btw 
    scheduler=lms,
    use_auth_token=True,
    cache_dir=os.getenv("cache_dir", "./models")
).to(device)


prompt = "a photo of an astronaut riding a horse on mars"
samples = 2
steps = 45
scale = 7.5
image = pipe(prompt)["sample"][0]
    

it definitely works, very slowly though.

On an M1 (not M1 Max) I get TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead. if I don't specify revision and torch_dtype. The script python scripts/txt2img.py works to create images though, so it's an issue with diffusers and not stable-diffusion I think.

On an M1 (not M1 Max) I get TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead. if I don't specify revision and torch_dtype. The script python scripts/txt2img.py works to create images though, so it's an issue with diffusers and not stable-diffusion I think.

Can you say how you specify revision and torch_dtype?

I think txt2image.py is using CPU if Cuda is not available

To sgt101, I think you're running in CPU mode, because of the line that says device = 'cuda' if torch.cuda.is_available() else 'cpu'

I'm pretty sure my txt2image is using MPS (magnusviri's fork) because it takes about a minute to run instead of upwards of 30 mins, among other things. I have

pipe = StableDiffusionPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4", 
    torch_dtype=torch.float16, revision="fp16",
    use_auth_token=True,
).to("mps")

But that errors out with

0it [00:00, ?it/s]loc("mps_add"("(mpsFileLoc): /AppleInternal/Library/BuildRoots/20d6c351-ee94-11ec-bcaf-7247572f23b4/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphUtilities.mm":219:0)): error: input types 'tensor<2x1280xf32>' and 'tensor<*xf16>' are not broadcast compatible
LLVM ERROR: Failed to infer result type(s).
Abort trap: 6
/Users/fragmede/miniforge3/envs/ldm/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

Figured it out! I opened a PR so hugging face can get my fix to diffusers to get MPS to work.

Figured it out! I opened a PR so hugging face can get my fix to diffusers to get MPS to work.

It works after upgrading diffusers.

 pip install -U diffusers

torch_dtype=torch.float16
remove this, and it works for me

Hey !

I have the same problem :

loc("varianceEps"("(mpsFileLoc): /AppleInternal/Library/BuildRoots/a0876c02-1788-11ed-b9c4-96898e02b808/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphUtilities.mm":219:0)): error: input types 'tensor<1x77x1xf16>' and 'tensor<1xf32>' are not broadcast compatible
LLVM ERROR: Failed to infer result type(s).

But I don't understand how to fix it ... I'm an architect and I don't have skills with coding ... Can you please develop a bit the method (if there is any) ?

Thanks a lot in advance !