Skip to content
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.

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

Protect alphas_cumprod during refiner switchover #14979

Merged
merged 5 commits into from
Mar 2, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
28 changes: 1 addition & 27 deletions modules/processing.py
Original file line number Diff line number Diff line change
Expand Up @@ -915,33 +915,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
if p.n_iter > 1:
shared.state.job = f"Batch {n+1} out of {p.n_iter}"

def rescale_zero_terminal_snr_abar(alphas_cumprod):
alphas_bar_sqrt = alphas_cumprod.sqrt()

# Store old values.
alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

# Shift so the last timestep is zero.
alphas_bar_sqrt -= (alphas_bar_sqrt_T)

# Scale so the first timestep is back to the old value.
alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

# Convert alphas_bar_sqrt to betas
alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
alphas_bar[-1] = 4.8973451890853435e-08
return alphas_bar

if hasattr(p.sd_model, 'alphas_cumprod') and hasattr(p.sd_model, 'alphas_cumprod_original'):
p.sd_model.alphas_cumprod = p.sd_model.alphas_cumprod_original.to(shared.device)

if opts.use_downcasted_alpha_bar:
p.extra_generation_params['Downcast alphas_cumprod'] = opts.use_downcasted_alpha_bar
p.sd_model.alphas_cumprod = p.sd_model.alphas_cumprod.half().to(shared.device)
if opts.sd_noise_schedule == "Zero Terminal SNR":
p.extra_generation_params['Noise Schedule'] = opts.sd_noise_schedule
p.sd_model.alphas_cumprod = rescale_zero_terminal_snr_abar(p.sd_model.alphas_cumprod).to(shared.device)
sd_models.apply_alpha_schedule_override(p.sd_model, p)

with devices.without_autocast() if devices.unet_needs_upcast else devices.autocast():
samples_ddim = p.sample(conditioning=p.c, unconditional_conditioning=p.uc, seeds=p.seeds, subseeds=p.subseeds, subseed_strength=p.subseed_strength, prompts=p.prompts)
Expand Down
32 changes: 32 additions & 0 deletions modules/sd_models.py
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@

from modules import paths, shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors, hashes, sd_models_config, sd_unet, sd_models_xl, cache, extra_networks, processing, lowvram, sd_hijack, patches
from modules.timer import Timer
from modules.shared import opts
import tomesd
import numpy as np

Expand Down Expand Up @@ -549,6 +550,36 @@ def repair_config(sd_config):
karlo_path = os.path.join(paths.models_path, 'karlo')
sd_config.model.params.noise_aug_config.params.clip_stats_path = sd_config.model.params.noise_aug_config.params.clip_stats_path.replace("checkpoints/karlo_models", karlo_path)

def apply_alpha_schedule_override(sd_model, p=None):
def rescale_zero_terminal_snr_abar(alphas_cumprod):
alphas_bar_sqrt = alphas_cumprod.sqrt()

# Store old values.
alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

# Shift so the last timestep is zero.
alphas_bar_sqrt -= (alphas_bar_sqrt_T)

# Scale so the first timestep is back to the old value.
alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

# Convert alphas_bar_sqrt to betas
alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
alphas_bar[-1] = 4.8973451890853435e-08
return alphas_bar

if hasattr(sd_model, 'alphas_cumprod') and hasattr(sd_model, 'alphas_cumprod_original'):
sd_model.alphas_cumprod = sd_model.alphas_cumprod_original.to(shared.device)

if opts.use_downcasted_alpha_bar:
if p is not None:
p.extra_generation_params['Downcast alphas_cumprod'] = opts.use_downcasted_alpha_bar
sd_model.alphas_cumprod = sd_model.alphas_cumprod.half().to(shared.device)
if opts.sd_noise_schedule == "Zero Terminal SNR":
if p is not None:
p.extra_generation_params['Noise Schedule'] = opts.sd_noise_schedule
sd_model.alphas_cumprod = rescale_zero_terminal_snr_abar(sd_model.alphas_cumprod).to(shared.device)

sd1_clip_weight = 'cond_stage_model.transformer.text_model.embeddings.token_embedding.weight'
sd2_clip_weight = 'cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight'
Expand Down Expand Up @@ -812,6 +843,7 @@ def reload_model_weights(sd_model=None, info=None, forced_reload=False):

sd_model = reuse_model_from_already_loaded(sd_model, checkpoint_info, timer)
if not forced_reload and sd_model is not None and sd_model.sd_checkpoint_info.filename == checkpoint_info.filename:
apply_alpha_schedule_override(sd_model)
return sd_model

if sd_model is not None:
Expand Down
Loading