ostris/ai-toolkit
1
0
Fork 0
mirror of https://github.com/ostris/ai-toolkit.git synced 2026-04-27 09:44:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use gradient checkpointing on DFE models if set

Browse Source
This commit is contained in:
Jaret Burkett
2025-04-11 10:45:39 -06:00
parent 4a43589666
commit cd37ccfc2e
1 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
extensions_built_in/sd_trainer/SDTrainer.py
View File

@@ -198,11 +198,21 @@ class SDTrainer(BaseSDTrainProcess):
if self.train_config.diffusion_feature_extractor_path is not None: if self.train_config.diffusion_feature_extractor_path is not None:
vae = None vae = None
if self.model_config.arch != "flux" or self.sd.vae.__class__.__name__ == "AutoencoderPixelMixer": if not (self.model_config.arch in ["flux"]) or self.sd.vae.__class__.__name__ == "AutoencoderPixelMixer":
vae = self.sd.vae vae = self.sd.vae
self.dfe = load_dfe(self.train_config.diffusion_feature_extractor_path, vae=vae) self.dfe = load_dfe(self.train_config.diffusion_feature_extractor_path, vae=vae)
self.dfe.to(self.device_torch) self.dfe.to(self.device_torch)
self.dfe.eval() if hasattr(self.dfe, 'vision_encoder') and self.train_config.gradient_checkpointing:
# must be set to train for gradient checkpointing to work
self.dfe.vision_encoder.train()
self.dfe.vision_encoder.gradient_checkpointing = True
else:
self.dfe.eval()
# enable gradient checkpointing on the vae
if vae is not None and self.train_config.gradient_checkpointing:
vae.enable_gradient_checkpointing()
vae.train()
def process_output_for_turbo(self, pred, noisy_latents, timesteps, noise, batch): def process_output_for_turbo(self, pred, noisy_latents, timesteps, noise, batch):