Upload pipeline.py

pipeline.py

@@ -1611,18 +1611,33 @@ class CustomPipeline(DiffusionPipeline, SD3LoraLoaderMixin):
  prompt_mask_input = prompt_mask
  latent_model_input = latents
 
+ if guidance_scale_real > 1.0 and i >= no_cfg_until_timestep:
+ progress_bar.set_postfix(
+ {
+ 'ts': timestep.detach().item() / 1000,
+ 'cfg': self._guidance_scale_real,
+ },
+ )
+ else:
+ progress_bar.set_postfix(
+ {
+ 'ts': timestep.detach().item() / 1000,
+ 'cfg': 'N/A',
+ },
+ )
+
  if do_batch_cfg and guidance_scale_real > 1.0 and i >= no_cfg_until_timestep:
  # Concatenate prompt embeddings
  prompt_embeds_input = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
  pooled_prompt_embeds_input = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
  
- # # Concatenate text IDs if they are used
- if text_ids is not None and negative_text_ids is not None:
- text_ids_input = torch.cat([negative_text_ids, text_ids], dim=0)
+ # Concatenate text IDs if they are used
+ # if text_ids is not None and negative_text_ids is not None:
+ # text_ids_input = torch.cat([negative_text_ids, text_ids], dim=0)
  
  # Concatenate latent image IDs if they are used
- if latent_image_ids is not None:
- latent_image_ids_input = torch.cat([latent_image_ids, latent_image_ids], dim=0)
+ # if latent_image_ids is not None:
+ # latent_image_ids_input = torch.cat([latent_image_ids, latent_image_ids], dim=0)
  
  # Concatenate prompt masks if they are used
  if prompt_mask is not None and negative_mask is not None:
@@ -1643,37 +1658,22 @@ class CustomPipeline(DiffusionPipeline, SD3LoraLoaderMixin):
  # Prepare extra transformer arguments
  extra_transformer_args = {}
  if prompt_mask is not None:
- extra_transformer_args["attention_mask"] = prompt_mask_input.to(device=self.transformer.device)
+ extra_transformer_args["attention_mask"] = prompt_mask_input.to(device=self.transformer.device).contiguous()
 
  # Forward pass through the transformer
  noise_pred = self.transformer(
- hidden_states=latent_model_input.to(device=self.transformer.device),
+ hidden_states=latent_model_input.to(device=self.transformer.device).contiguous() ,
  timestep=timestep / 1000,
  guidance=guidance,
- pooled_projections=pooled_prompt_embeds_input.to(device=self.transformer.device),
- encoder_hidden_states=prompt_embeds_input.to(device=self.transformer.device),
- txt_ids=text_ids_input.to(device=self.transformer.device) if text_ids is not None else None,
- img_ids=latent_image_ids_input.to(device=self.transformer.device) if latent_image_ids is not None else None,
+ pooled_projections=pooled_prompt_embeds_input.to(device=self.transformer.device).contiguous() ,
+ encoder_hidden_states=prompt_embeds_input.to(device=self.transformer.device).contiguous() ,
+ txt_ids=text_ids_input.to(device=self.transformer.device).contiguous() if text_ids is not None else None,
+ img_ids=latent_image_ids_input.to(device=self.transformer.device).contiguous() if latent_image_ids is not None else None,
  joint_attention_kwargs=self.joint_attention_kwargs,
  return_dict=False,
  **extra_transformer_args,
  )[0]
 
- if do_batch_cfg and guidance_scale_real > 1.0 and i >= no_cfg_until_timestep:
- progress_bar.set_postfix(
- {
- 'ts': timestep.detach().item() / 1000,
- 'cfg': self._guidance_scale_real,
- },
- )
- else:
- progress_bar.set_postfix(
- {
- 'ts': timestep.detach().item() / 1000,
- 'cfg': 'N/A',
- },
- )
-
  # Apply real CFG
  if guidance_scale_real > 1.0 and i >= no_cfg_until_timestep:
  if do_batch_cfg: