[Qwen LoRA training] fix bug when offloading (#12440)

* fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled

[Qwen LoRA training] fix bug when offloading (#12440)
* fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled * fix bug when offload and cache_latents both enabled
1066de8c · Linoy Tsaban · GitHub · 2d69bacb · 1066de8c
Unverified Commit 1066de8c authored Oct 07, 2025 by Linoy Tsaban Committed by GitHub Oct 07, 2025
Hide whitespace changes
Inline Side-by-side

Showing with 1 addition and 1 deletion

examples/dreambooth/train_dreambooth_lora_qwen_image.py examples/dreambooth/train_dreambooth_lora_qwen_image.py +1 -1

No files found.
--- a/examples/dreambooth/train_dreambooth_lora_qwen_image.py
+++ b/examples/dreambooth/train_dreambooth_lora_qwen_image.py
@@ -1338,7 +1338,7 @@ def main(args):
                        batch["pixel_values"] = batch["pixel_values"].to(
                            accelerator.device, non_blocking=True, dtype=vae.dtype
                        )
-                    latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
+                        latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
                if train_dataset.custom_instance_prompts:
                    with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                        prompt_embeds, prompt_embeds_mask = compute_text_embeddings(