Update

1 files changed, 11 insertions, 14 deletions

diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index ccec215..cab5e4c 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -11,10 +11,7 @@ from transformers import CLIPTextModel
 from diffusers import AutoencoderKL, UNet2DConditionModel, DPMSolverMultistepScheduler
 from diffusers.loaders import AttnProcsLayers
 
-from slugify import slugify
-
 from models.clip.tokenizer import MultiCLIPTokenizer
-from training.util import EMAModel
 from training.functional import TrainingStrategy, TrainingCallbacks, save_samples
 
 
@@ -41,16 +38,9 @@ def lora_strategy_callbacks(
     sample_output_dir.mkdir(parents=True, exist_ok=True)
     checkpoint_output_dir.mkdir(parents=True, exist_ok=True)
 
-    weight_dtype = torch.float32
-    if accelerator.state.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif accelerator.state.mixed_precision == "bf16":
-        weight_dtype = torch.bfloat16
-
     save_samples_ = partial(
         save_samples,
         accelerator=accelerator,
-        unet=unet,
         text_encoder=text_encoder,
         tokenizer=tokenizer,
         vae=vae,
@@ -83,20 +73,27 @@ def lora_strategy_callbacks(
         yield
 
     def on_before_optimize(lr: float, epoch: int):
-        if accelerator.sync_gradients:
-            accelerator.clip_grad_norm_(lora_layers.parameters(), max_grad_norm)
+        accelerator.clip_grad_norm_(lora_layers.parameters(), max_grad_norm)
 
     @torch.no_grad()
     def on_checkpoint(step, postfix):
         print(f"Saving checkpoint for step {step}...")
 
         unet_ = accelerator.unwrap_model(unet, False)
-        unet_.save_attn_procs(checkpoint_output_dir / f"{step}_{postfix}")
+        unet_.save_attn_procs(
+            checkpoint_output_dir / f"{step}_{postfix}",
+            safe_serialization=True
+        )
         del unet_
 
     @torch.no_grad()
     def on_sample(step):
-        save_samples_(step=step)
+        unet_ = accelerator.unwrap_model(unet, False)
+        save_samples_(step=step, unet=unet_)
+        del unet_
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
 
     return TrainingCallbacks(
         on_prepare=on_prepare,