update

Author: westfish

ppdiffusers/examples/dreambooth/README_sd3.md

@@ -48,14 +48,14 @@ python train_dreambooth_sd3.py \
   --report_to="wandb" \
   --lr_scheduler="constant" \
   --lr_warmup_steps=0 \
-  --max_train_steps=500 \
+  --max_train_steps=50 \
   --validation_prompt="A photo of sks dog in a bucket" \
   --validation_epochs=25 \
   --seed="0" \
   --checkpointing_steps=250
 ```
 
-为了更好地跟踪我们的训练实验，我们在上面的命令中使用了以下标志：
+fp16训练需要显存47000MiB，为了更好地跟踪我们的训练实验，我们在上面的命令中使用了以下标志：
 
 * `report_to="wandb"` 将确保在 Weights and Biases 上跟踪训练运行。要使用它，请确保安装 `wandb`，使用 `pip install wandb`。
 * `validation_prompt` 和 `validation_epochs` 允许脚本进行几次验证推理运行。这可以让我们定性地检查训练是否按预期进行。

ppdiffusers/examples/dreambooth/requirements_sd3.txt

@@ -1,3 +1,2 @@
 Pillow
-ppdiffusers>=0.24.1
 wandb

ppdiffusers/examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1510,6 +1510,9 @@ def get_state_dict(model):
                 del text_encoder_one, text_encoder_two, text_encoder_three
                 gc.collect()
                 paddle.device.cuda.empty_cache()
+                
+        if global_step >= args.max_train_steps:
+            break
 
     # Save the lora layers
     accelerator.wait_for_everyone()

ppdiffusers/examples/dreambooth/train_dreambooth_sd3.py

@@ -1578,7 +1578,9 @@ def get_state_dict(model):
                     gc.collect()
                     paddle.device.cuda.empty_cache()
 
-    # Save the lora layers
+        if global_step >= args.max_train_steps:
+            break
+
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         transformer = unwrap_model(transformer)

ppdiffusers/ppdiffusers/loaders/lora.py

@@ -1483,7 +1483,7 @@ class SD3LoraLoaderMixin:
     num_fused_loras = 0
 
     def load_lora_weights(
-        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, paddle.Tensor]], adapter_name=None, **kwargs
     ):
         """
         Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
@@ -1781,7 +1781,7 @@ def pack_weights(layers, prefix):
 
     @staticmethod
     def write_lora_layers(
-        state_dict: Dict[str, torch.Tensor],
+        state_dict: Dict[str, paddle.Tensor],
         save_directory: str,
         is_main_process: bool,
         weight_name: str,