indices should be either on cpu or on the same device as the indexed tensor (cpu)

[andydhancock]

Describe the bug

image_to_image.py line 92 throws the error above.

  `  init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)`

I've tried adding .to(self.device) to the 3 parameters.
Device should be 'cuda' though.

Reproduction

device = "cuda"

pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16,
    use_auth_token=True
).to(device)

response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')

init_image = Image.open(io.BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))
init_image = preprocess(init_image)

outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)

Logs

No response

System Info

diffusers==0.2.4
nvidia/cuda:11.3.0-cudnn8-devel-ubuntu20.04

[patrickvonplaten]

@patil-suraj could you take a look here?

[andydhancock]

I downloaded the source and tracked the problem down to this line...
diffusers/schedulers/scheduling_pndm.py:254

    sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5

[andydhancock]

Solved:
Maybe I should listen to the error message sometime.. it should all be cpu
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
becomes
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps.cpu())
and all is well.

[loretoparisi]

Solved:

Maybe I should listen to the error message sometime.. it should all be cpu

init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)

becomes

init_latents = self.scheduler.add_noise(init_latents, noise, timesteps.cpu())

and all is well.

Is .to(device) not supported for timesteps then?

[andydhancock]

Solved:
Maybe I should listen to the error message sometime.. it should all be cpu
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
becomes
init_latents = self.scheduler.add_noise(init_latents, noise, timesteps.cpu())
and all is well.

Is .to(device) not supported for timesteps then?

I think it was supported but 'device' was cuda, and it's trying to do something with numpy which uses cpu

[patil-suraj]

Hi @andydhancock

This is working fine for me. Here's the code snippet I tried.

cd into examples/inference and run

import requests
from PIL import Image
from io import BytesIO

from torch import autocast


from image_to_image import *

device = "cuda"
pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16,
    use_auth_token=True
).to(device)


response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))
init_image = preprocess(init_image)

prompt = "a cat, artstation"

with autocast("cuda"):
    outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)

[andydhancock]

Hi @andydhancock

This is working fine for me. Here's the code snippet I tried.

cd into examples/inference and run

import requests
from PIL import Image
from io import BytesIO

from torch import autocast


from image_to_image import *

device = "cuda"
pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16,
    use_auth_token=True
).to(device)


response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))
init_image = preprocess(init_image)

prompt = "a cat, artstation"

with autocast("cuda"):
    outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)

Strange it wouldn't on mine then! I got it to work with the timesteps.cpu() thing anyways so not a problem anymore.

[loretoparisi]

@andydhancock thanks! On Apple M1 it worked in cpu removing FP16 and autocast:
🥇

import requests
from PIL import Image
from io import BytesIO

from torch import autocast

from image_to_image import *

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

pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    #revision="fp16", 
    #torch_dtype=torch.float16,
    use_auth_token=True
).to(device)


response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))
init_image = preprocess(init_image)

prompt = "a cat, artstation"
samples = 1
outputs = []
if device=='cuda':
    with autocast("cuda"):
        outputs = pipei2i(prompt=[prompt]*samples, 
            init_image=init_image, 
            #strength=0.75, 
            num_inference_steps=50,
            guidance_scale=7.5)
else:
    outputs = pipei2i(prompt=[prompt]*samples,
        init_image=init_image, 
        #strength=0.75, 
        num_inference_steps=50,
        guidance_scale=7.5)

A question, the default value guidance_scale was 0.75 not 7.5 as above, right? And same for strength I assume.
Thank you.

[andydhancock]

Yeah, guidance_scale is 7.5, strength 0.75.. typo.. they're variables in my actual code

[ghost]

Wait why was this closed? Shouldn't this be fixed in source? Or is the Collab someone else's problem?

[patrickvonplaten]

Doesn't this work for you @cherrerajobs:

import requests
from PIL import Image
from io import BytesIO

from torch import autocast


from diffusers import StableDiffusionImg2ImgPipeline

device = "cuda"
pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "CompVis/stable-diffusion-v1-4",
    revision="fp16", 
    torch_dtype=torch.float16,
    use_auth_token=True
).to(device)


response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))

prompt = "a cat, artstation"

with autocast("cuda"):
    outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)

It should work now on master

[FahimF]

Using the code from main branch and the following code, it fails on Apple Silicon devices:

import torch
import requests
from PIL import Image
from io import BytesIO

from torch import autocast


from diffusers import StableDiffusionImg2ImgPipeline

device = torch.device("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu")
pipei2i = StableDiffusionImg2ImgPipeline.from_pretrained(
    "stable-diffusion-v1-4",
).to(device)


response = requests.get('https://pbs.twimg.com/media/Fa1_7_vWYAEwfX-.png')
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))

prompt = "a cat, artstation"

with autocast("cpu"):
    outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)

The error is as follows:

Traceback (most recent call last):
  File "test.py", line 24, in <module>
    outputs = pipei2i(prompt=prompt, init_image=init_image, strength=0.75, num_inference_steps=75,guidance_scale=0.75)
  File "/Users/fahim/miniforge3/envs/ml/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "/Users/fahim/miniforge3/envs/ml/lib/python3.8/site-packages/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py", line 101, in __call__
    init_latents = self.scheduler.add_noise(init_latents, noise, timesteps)
  File "/Users/fahim/miniforge3/envs/ml/lib/python3.8/site-packages/diffusers/schedulers/scheduling_pndm.py", line 270, in add_noise
    sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu)

It appears that self.alphas is on cpu while timesteps is on mps in scheduling_pndm.py on line 268? (I know it says 270 but I made some changes to the code there to check something ...)

If I change references to timesteps to timesteps.cpu() in that method, then the code works. But wouldn't it be better to put self.alphas_cumprod on whatever device is being used and so basically put that on the current device instead? Just asking since I'm not sure if that is a better idea or not ...

[patrickvonplaten]

Interesting! I don't have a mac computer - @patil-suraj @pcuenca do you have one? Could you give it a try maybe? :-)

[FahimF]

@patrickvonplaten Thanks for looking into this. I'm able to get around this currently by simply modifying code to do the following instead of the existing code at line 268 in scheduling_pndm.py:

        device = torch.device("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu")
        arr = self.alphas_cumprod.to(device)

Then I usee arr instead of self.alphas_cumprod in the rest of the method. Just wasn't sure if that was the best way to go ... I'm a developer, but not a Python developer 😄 So if there's anything I can do to help with troubleshooting the issue, please do let me know.

[patrickvonplaten]

Think @pcuenca is currently working on a fix in #355 🙂

[FahimF]

Thanks guys 😄 Is there anything I should have done differently about this bug report? Just asking in case I should have reported MPS issues elsewhere?

I do have two other issues (both on MPS) which are not crashes but the behaviour is different to what is expected (and what is observed on non-MPS devices). So just wondering if I should report or wait till you guys are done with the MPS changes first?

[pcuenca]

@FahimF Thanks for reporting this!

By all means, do report any other issues you've found. If you are sure they are related with the mps device, feel free to say so in the title :)

[FahimF]

@pcuenca Thanks, will do 🙂 And yes, both are mps device issues though I don't know if they are in the Hugging Face code or upstream in Pytorch. Just that the behaviour is distinctly different than on a non-mps device. But will open new issues for those rather than discussing here ... unless there is already an existing issue. Haven't checked yet ...

[FahimF]

This issue appears to have cropped up again in the 0.3.0 release in scheduling_lms_discrete.py at line 187. I believe the fix might have been made in scheduling_pndm.py only perhaps?

Update: tagging @pcuenca since the ticket is closed and not sure if anybody gets notified.

[patrickvonplaten]

Should we maybe open a new issue here stating that this concerns mostly "mps"?

[pcuenca]

I just opened #501. If we keep discovering mps issues we could create a label for that.