Skip to content

Proof of concept fix for Flax LCM PR #1

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

Open
wants to merge 2 commits into
base: lcm_flax
Choose a base branch
from
Open

Proof of concept fix for Flax LCM PR #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7e0a0cd
Allow diffusers to load with Flax, w/o PyTorch
pcuenca Dec 21, 2023
2102411
Split key in denoising loop
pcuenca Jan 6, 2024
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
9 changes: 5 additions & 4 deletions src/diffusers/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,7 +230,7 @@ def _generate(

# Denoising loop
def loop_body(step, args):
latents, scheduler_state = args
latents, scheduler_state, prng_seed = args
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
Expand All @@ -255,14 +255,15 @@ def loop_body(step, args):

# compute the previous noisy sample x_t -> x_t-1
latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents, prng_seed).to_tuple()
return latents, scheduler_state
prng_seed = jax.random.split(prng_seed)[0]
return latents, scheduler_state, prng_seed

if DEBUG:
# run with python for loop
for i in range(num_inference_steps):
latents, scheduler_state = loop_body(i, (latents, scheduler_state))
latents, scheduler_state, prng_seed = loop_body(i, (latents, scheduler_state, prng_seed))
else:
latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state))
latents, _, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state, prng_seed))

if return_latents:
return latents
Expand Down
6 changes: 3 additions & 3 deletions src/diffusers/utils/torch_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,7 +89,7 @@ def is_compiled_module(module) -> bool:
return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)


def fourier_filter(x_in: torch.Tensor, threshold: int, scale: int) -> torch.Tensor:
def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
"""Fourier filter as introduced in FreeU (https://arxiv.org/abs/2309.11497).

This version of the method comes from here:
Expand Down Expand Up @@ -121,8 +121,8 @@ def fourier_filter(x_in: torch.Tensor, threshold: int, scale: int) -> torch.Tens


def apply_freeu(
resolution_idx: int, hidden_states: torch.Tensor, res_hidden_states: torch.Tensor, **freeu_kwargs
) -> Tuple[torch.Tensor, torch.Tensor]:
resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs
) -> Tuple["torch.Tensor", "torch.Tensor"]:
"""Applies the FreeU mechanism as introduced in https:
//arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU.

Expand Down