Fixed Lora training

4 files changed, 35 insertions, 37 deletions

diff --git a/train_dreambooth.py b/train_dreambooth.py
index 5a4c47b..a29c507 100644
--- a/train_dreambooth.py
+++ b/train_dreambooth.py
@@ -442,6 +442,12 @@ def main():
         mixed_precision=args.mixed_precision
     )
 
+    weight_dtype = torch.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
     logging.basicConfig(filename=output_dir.joinpath("log.txt"), level=logging.DEBUG)
 
     if args.seed is None:
@@ -495,12 +501,6 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
-    weight_dtype = torch.float32
-    if args.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif args.mixed_precision == "bf16":
-        weight_dtype = torch.bfloat16
-
     trainer = partial(
         train,
         accelerator=accelerator,
diff --git a/train_lora.py b/train_lora.py
index b273ae1..ab1753b 100644
--- a/train_lora.py
+++ b/train_lora.py
@@ -13,7 +13,7 @@ from accelerate.logging import get_logger
 from accelerate.utils import LoggerType, set_seed
 from slugify import slugify
 from diffusers.loaders import AttnProcsLayers
-from diffusers.models.cross_attention import LoRAXFormersCrossAttnProcessor
+from diffusers.models.cross_attention import LoRAXFormersCrossAttnProcessor, LoRACrossAttnProcessor
 
 from util import load_config, load_embeddings_from_dir
 from data.csv import VlpnDataModule, keyword_filter
@@ -178,6 +178,11 @@ def parse_args():
         help="Number of updates steps to accumulate before performing a backward/update pass.",
     )
     parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
         "--find_lr",
         action="store_true",
         help="Automatically find a learning rate (no training).",
@@ -402,6 +407,12 @@ def main():
         mixed_precision=args.mixed_precision
     )
 
+    weight_dtype = torch.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
     logging.basicConfig(filename=output_dir.joinpath("log.txt"), level=logging.DEBUG)
 
     if args.seed is None:
@@ -418,6 +429,12 @@ def main():
     vae.set_use_memory_efficient_attention_xformers(True)
     unet.enable_xformers_memory_efficient_attention()
 
+    if args.gradient_checkpointing:
+        unet.enable_gradient_checkpointing()
+
+    unet.to(accelerator.device, dtype=weight_dtype)
+    text_encoder.to(accelerator.device, dtype=weight_dtype)
+
     lora_attn_procs = {}
     for name in unet.attn_processors.keys():
         cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
@@ -467,12 +484,6 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
-    weight_dtype = torch.float32
-    if args.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif args.mixed_precision == "bf16":
-        weight_dtype = torch.bfloat16
-
     trainer = partial(
         train,
         accelerator=accelerator,
diff --git a/train_ti.py b/train_ti.py
index 56f9e97..2840def 100644
--- a/train_ti.py
+++ b/train_ti.py
@@ -513,6 +513,12 @@ def main():
         mixed_precision=args.mixed_precision
     )
 
+    weight_dtype = torch.float32
+    if args.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif args.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
     logging.basicConfig(filename=output_dir.joinpath("log.txt"), level=logging.DEBUG)
 
     if args.seed is None:
@@ -564,12 +570,6 @@ def main():
     else:
         optimizer_class = torch.optim.AdamW
 
-    weight_dtype = torch.float32
-    if args.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif args.mixed_precision == "bf16":
-        weight_dtype = torch.bfloat16
-
     checkpoint_output_dir = output_dir.joinpath("checkpoints")
 
     trainer = partial(
diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index 92abaa6..bc10e58 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -89,20 +89,14 @@ def lora_strategy_callbacks(
     @torch.no_grad()
     def on_checkpoint(step, postfix):
         print(f"Saving checkpoint for step {step}...")
-        orig_unet_dtype = unet.dtype
-        unet.to(dtype=torch.float32)
-        unet.save_attn_procs(checkpoint_output_dir.joinpath(f"{step}_{postfix}"))
-        unet.to(dtype=orig_unet_dtype)
+
+        unet_ = accelerator.unwrap_model(unet)
+        unet_.save_attn_procs(checkpoint_output_dir.joinpath(f"{step}_{postfix}"))
+        del unet_
 
     @torch.no_grad()
     def on_sample(step):
-        orig_unet_dtype = unet.dtype
-        unet.to(dtype=weight_dtype)
         save_samples_(step=step)
-        unet.to(dtype=orig_unet_dtype)
-
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
 
     return TrainingCallbacks(
         on_prepare=on_prepare,
@@ -126,16 +120,9 @@ def lora_prepare(
     lora_layers: AttnProcsLayers,
     **kwargs
 ):
-    weight_dtype = torch.float32
-    if accelerator.state.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif accelerator.state.mixed_precision == "bf16":
-        weight_dtype = torch.bfloat16
-
     lora_layers, optimizer, train_dataloader, val_dataloader, lr_scheduler = accelerator.prepare(
         lora_layers, optimizer, train_dataloader, val_dataloader, lr_scheduler)
-    unet.to(accelerator.device, dtype=weight_dtype)
-    text_encoder.to(accelerator.device, dtype=weight_dtype)
+
     return text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler, {"lora_layers": lora_layers}