[Flux Dreambooth lora] add latent caching

examples/dreambooth/README_flux.md

@@ -8,7 +8,7 @@ The `train_dreambooth_flux.py` script shows how to implement the training proced
 > 
 > Flux can be quite expensive to run on consumer hardware devices and as a result finetuning it comes with high memory requirements - 
 > a LoRA with a rank of 16 (w/ all components trained) can exceed 40GB of VRAM for training.
-> For more tips & guidance on training on a resource-constrained device please visit [`@bghira`'s guide](https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md) 
+> For more tips & guidance on training on a resource-constrained device please check out these great guides and trainers for FLUX: [`@bghira`'s guide](https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md) & [`ostris` guide](https://github.com/ostris/ai-toolkit?tab=readme-ov-file#flux1-training) 
 
 
 > [!NOTE]

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -599,6 +599,12 @@ def parse_args(input_args=None):
             " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
         ),
     )
+    parser.add_argument(
+        "--cache_latents",
+        action="store_true",
+        default=False,
+        help="Cache the VAE latents",
+    )
     parser.add_argument(
         "--report_to",
         type=str,
@@ -1456,6 +1462,24 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 tokens_one = torch.cat([tokens_one, class_tokens_one], dim=0)
                 tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)
 
+    vae_config_shift_factor = vae.config.shift_factor
+    vae_config_scaling_factor = vae.config.scaling_factor
+    vae_config_block_out_channels = vae.config.block_out_channels
+    if args.cache_latents:
+        latents_cache = []
+        for batch in tqdm(train_dataloader, desc="Caching latents"):
+            with torch.no_grad():
+                batch["pixel_values"] = batch["pixel_values"].to(
+                    accelerator.device, non_blocking=True, dtype=weight_dtype
+                )
+                latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
+
+        if args.validation_prompt is None:
+            del vae
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+                gc.collect()
+
     # Scheduler and math around the number of training steps.
     overrode_max_train_steps = False
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -1578,7 +1602,6 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
             if args.train_text_encoder:
                 models_to_accumulate.extend([text_encoder_one])
             with accelerator.accumulate(models_to_accumulate):
-                pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                 prompts = batch["prompts"]
 
                 # encode batch prompts when custom prompts are provided for each image -
@@ -1610,11 +1633,15 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                         )
 
                 # Convert images to latent space
-                model_input = vae.encode(pixel_values).latent_dist.sample()
-                model_input = (model_input - vae.config.shift_factor) * vae.config.scaling_factor
+                if args.cache_latents:
+                    model_input = latents_cache[step].sample()
+                else:
+                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
+                    model_input = vae.encode(pixel_values).latent_dist.sample()
+                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                 model_input = model_input.to(dtype=weight_dtype)
 
-                vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
+                vae_scale_factor = 2 ** (len(vae_config_block_out_channels))
 
                 latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                     model_input.shape[0],