[Bug]: Apple silicon can't use dpm++ 2m sampling model

[sung-ho-moon]

Is there an existing issue for this?

• I have searched the existing issues and checked the recent builds/commits

What happened?

when I use SD,
I can't use dpm++ 2m, dpm++ 2m karras sampler
All results is black screen.

Steps to reproduce the problem

when I use dpm++2m, dpm++2m karras
results is black picture

I'm using SD, Apple silicon M1

I editted K-diffusion/k-diffusion/sampling.py

//
def to_d(x, sigma, denoised):
"""Converts a denoiser output to a Karras ODE derivative."""
sigma = sigma.to('cpu').to('mps')
return (x - denoised) / utils.append_dims(sigma, x.ndim)
//

Is anyone using dpm++ 2m Apple silicon?

What should have happened?

get the picture

Commit where the problem happens

ac08562

What platforms do you use to access UI ?

No response

What browsers do you use to access the UI ?

No response

Command Line Arguments

No response

Additional information, context and logs

No response

[remixer-dec]

Same bug here.
Looks like the problem is that this function returns the same value for different inputs, this is really weird, because exactly the same function declared a few lines above it, behaves correctly.

anyways, here is a fix:

t_fn = lambda sigma: sigma.to('cpu').log().neg().to('mps')

Update: this bug only appears when list indices are used and is fixed in the latest pytorch nightly (but there are some known problems currently in the nightly builds, so I do not recommend upgrading)

[sung-ho-moon]

@remixer-dec thank you, problem solved

[remixer-dec]

@moon-haha
I think this fix should be also applied to DPM++ 2S a, it is currently working only because of noise, here is an example

[sung-ho-moon]

@remixer-dec
that's right, DPM++ 2S a Has Noise Bug too, again Thank you

[lsooxlla8]

anyways, here is a fix:

t_fn = lambda sigma: sigma.to('cpu').log().neg().to('mps')

Could you please explain where to paste it to, for noncoders?

[remixer-dec]

@lsooxlla8 repositories/k-diffusion/sampling.py replace this line and this line and make sure the indentation is the same

[modusCell]

@remixer-dec

    # t_fn = lambda sigma: sigma.log().neg()
    t_fn = lambda sigma: sigma.to('cpu').log().neg().to('mps')

Commented out that line and inserted the one you suggested, restarted ComfyUI and everything. it's still the same.
Let's say there are 8 steps, first 6 steps I can see image / noise, after finishing step 6 image turn to black.

This applies to SD, XL and Turbo for all DMP++

cli command for starting ComfyUI:

./venv/bin/python3 main.py --force-fp16 --dont-upcast-attention --disable-xformers