Skip to content

Fix MPS scheduler indexing when using mps #450

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Sep 14, 2022
Merged

Fix MPS scheduler indexing when using mps #450

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6a0c1db
Fix LMS scheduler indexing in `add_noise` #358.
pcuenca Sep 9, 2022
c8e6b26
Fix DDIM and DDPM indexing with mps device.
pcuenca Sep 9, 2022
bdaceed
Verify format is PyTorch before using `.to()`
pcuenca Sep 14, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions src/diffusers/schedulers/scheduling_ddim.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,6 +249,7 @@ def add_noise(
noise: Union[torch.FloatTensor, np.ndarray],
timesteps: Union[torch.IntTensor, np.ndarray],
) -> Union[torch.FloatTensor, np.ndarray]:
timesteps = timesteps.to(self.alphas_cumprod.device)
Copy link
Member

@anton-l anton-l Sep 13, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This breaks the assumption that we can work with numpy arrays as well :(
Maybe we could add a device argument to scheduler.set_format() and call if from pipeline.to()?

Copy link
Member Author

@pcuenca pcuenca Sep 13, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd rather minimize the changes across files just for this particular case that only affects mps. How about some defensive coding like:

  • Check the format is pt
  • And alphas_cumprod is actually a tensor

Then move using to() ?

Copy link
Contributor

@patrickvonplaten patrickvonplaten Sep 13, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Think we'll move away from set_format() since we'll soon have framework dependent schedulers - so I think this is ok here

sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
sqrt_alpha_prod = self.match_shape(sqrt_alpha_prod, original_samples)
sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
Expand Down
1 change: 1 addition & 0 deletions src/diffusers/schedulers/scheduling_ddpm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,6 +250,7 @@ def add_noise(
noise: Union[torch.FloatTensor, np.ndarray],
timesteps: Union[torch.IntTensor, np.ndarray],
) -> Union[torch.FloatTensor, np.ndarray]:
timesteps = timesteps.to(self.alphas_cumprod.device)
sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
sqrt_alpha_prod = self.match_shape(sqrt_alpha_prod, original_samples)
sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
Expand Down
3 changes: 2 additions & 1 deletion src/diffusers/schedulers/scheduling_lms_discrete.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,7 +121,7 @@ def set_timesteps(self, num_inference_steps: int):
frac = np.mod(self.timesteps, 1.0)
sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
sigmas = (1 - frac) * sigmas[low_idx] + frac * sigmas[high_idx]
self.sigmas = np.concatenate([sigmas, [0.0]])
self.sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)

self.derivatives = []

Expand Down Expand Up @@ -184,6 +184,7 @@ def add_noise(
noise: Union[torch.FloatTensor, np.ndarray],
timesteps: Union[torch.IntTensor, np.ndarray],
) -> Union[torch.FloatTensor, np.ndarray]:
timesteps = timesteps.to(self.sigmas.device)
sigmas = self.match_shape(self.sigmas[timesteps], noise)
noisy_samples = original_samples + noise * sigmas

Expand Down