[Flux Dreambooth lora] add latent caching (huggingface#9160)

* add ostris trainer to README & add cache latents of vae

* add ostris trainer to README & add cache latents of vae

* style

* readme

* add test for latent caching

* add ostris noise scheduler
https://github.com/ostris/ai-toolkit/blob/9ee1ef2a0a2a9a02b92d114a95f21312e5906e54/toolkit/samplers/custom_flowmatch_sampler.py#L95

* style

* fix import

* style

* fix tests

* style

* --change upcasting of transformer?

* update readme according to main

* keep only latent caching

* add configurable param for final saving of trained layers- --upcast_before_saving

* style

* Update examples/dreambooth/README_flux.md

Co-authored-by: Sayak Paul <[email protected]>

* Update examples/dreambooth/README_flux.md

Co-authored-by: Sayak Paul <[email protected]>

* use clear_objs_and_retain_memory from utilities

* style

---------

Co-authored-by: Sayak Paul <[email protected]>

Author: linoytsaban

examples/dreambooth/README_flux.md

@@ -221,8 +221,12 @@ Instead, only a subset of these activations (the checkpoints) are stored and the
 ### 8-bit-Adam Optimizer
 When training with `AdamW`(doesn't apply to `prodigy`) You can pass `--use_8bit_adam` to reduce the memory requirements of training. 
 Make sure to install `bitsandbytes` if you want to do so.
-### latent caching
+### Latent caching
 When training w/o validation runs, we can pre-encode the training images with the vae, and then delete it to free up some memory. 
-to enable `latent_caching`, first, use the version in [this PR](https://github.com/huggingface/diffusers/blob/1b195933d04e4c8281a2634128c0d2d380893f73/examples/dreambooth/train_dreambooth_lora_flux.py), and then pass `--cache_latents`
+to enable `latent_caching` simply pass `--cache_latents`.
+### Precision of saved LoRA layers
+By default, trained transformer layers are saved in the precision dtype in which training was performed. E.g. when training in mixed precision is enabled with `--mixed_precision="bf16"`, final finetuned layers will be saved in `torch.bfloat16` as well. 
+This reduces memory requirements significantly w/o a significant quality loss. Note that if you do wish to save the final layers in float32 at the expanse of more memory usage, you can do so by passing `--upcast_before_saving`.
+
 ## Other notes
 Thanks to `bghira` and `ostris` for their help with reviewing & insight sharing ♥️

examples/dreambooth/test_dreambooth_lora_flux.py

@@ -103,6 +103,39 @@ def test_dreambooth_lora_text_encoder_flux(self):
             )
             self.assertTrue(starts_with_expected_prefix)
 
+    def test_dreambooth_lora_latent_caching(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+                {self.script_path}
+                --pretrained_model_name_or_path {self.pretrained_model_name_or_path}
+                --instance_data_dir {self.instance_data_dir}
+                --instance_prompt {self.instance_prompt}
+                --resolution 64
+                --train_batch_size 1
+                --gradient_accumulation_steps 1
+                --max_train_steps 2
+                --cache_latents
+                --learning_rate 5.0e-04
+                --scale_lr
+                --lr_scheduler constant
+                --lr_warmup_steps 0
+                --output_dir {tmpdir}
+                """.split()
+
+            run_command(self._launch_args + test_args)
+            # save_pretrained smoke test
+            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))
+
+            # make sure the state_dict has the correct naming in the parameters.
+            lora_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
+            is_lora = all("lora" in k for k in lora_state_dict.keys())
+            self.assertTrue(is_lora)
+
+            # when not training the text encoder, all the parameters in the state dict should start
+            # with `"transformer"` in their names.
+            starts_with_transformer = all(key.startswith("transformer") for key in lora_state_dict.keys())
+            self.assertTrue(starts_with_transformer)
+
     def test_dreambooth_lora_flux_checkpointing_checkpoints_total_limit(self):
         with tempfile.TemporaryDirectory() as tmpdir:
             test_args = f"""

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -15,7 +15,6 @@
 
 import argparse
 import copy
-import gc
 import itertools
 import logging
 import math
@@ -56,6 +55,7 @@
 from diffusers.training_utils import (
     _set_state_dict_into_text_encoder,
     cast_training_params,
+    clear_objs_and_retain_memory,
     compute_density_for_timestep_sampling,
     compute_loss_weighting_for_sd3,
 )
@@ -600,6 +600,12 @@ 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,
@@ -620,6 +626,15 @@ def parse_args(input_args=None):
             " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
         ),
     )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
     parser.add_argument(
         "--prior_generation_precision",
         type=str,
@@ -1422,12 +1437,7 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
 
     # Clear the memory here
     if not args.train_text_encoder and not train_dataset.custom_instance_prompts:
-        del tokenizers, text_encoders
-        # Explicitly delete the objects as well, otherwise only the lists are deleted and the original references remain, preventing garbage collection
-        del text_encoder_one, text_encoder_two
-        gc.collect()
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
+        clear_objs_and_retain_memory([tokenizers, text_encoders, text_encoder_one, text_encoder_two])
 
     # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
     # pack the statically computed variables appropriately here. This is so that we don't
@@ -1457,6 +1467,21 @@ 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:
+            clear_objs_and_retain_memory([vae])
+
     # 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)
@@ -1579,7 +1604,6 @@ 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 -
@@ -1613,11 +1637,15 @@ 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],
@@ -1789,15 +1817,16 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                     torch_dtype=weight_dtype,
                 )
                 if not args.train_text_encoder:
-                    del text_encoder_one, text_encoder_two
-                    torch.cuda.empty_cache()
-                    gc.collect()
+                    clear_objs_and_retain_memory([text_encoder_one, text_encoder_two])
 
     # Save the lora layers
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         transformer = unwrap_model(transformer)
-        transformer = transformer.to(torch.float32)
+        if args.upcast_before_saving:
+            transformer.to(torch.float32)
+        else:
+            transformer = transformer.to(weight_dtype)
         transformer_lora_layers = get_peft_model_state_dict(transformer)
 
         if args.train_text_encoder: