Fix InstructPix2Pix training in multi-GPU mode (huggingface#2978)

* fix: norm group test for UNet3D.

* fix: unet rejig.

* fix: unwrapping when running validation inputs.

* unwrapping the unet too.

* fix: device.

* better unwrapping.

* unwrapping before ema.

* unwrapping.

Author: sayakpaul

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -451,19 +451,18 @@ def main():
     # then fine-tuned on the custom InstructPix2Pix dataset. This modified UNet is initialized
     # from the pre-trained checkpoints. For the extra channels added to the first layer, they are
     # initialized to zero.
-    if accelerator.is_main_process:
-        logger.info("Initializing the InstructPix2Pix UNet from the pretrained UNet.")
-        in_channels = 8
-        out_channels = unet.conv_in.out_channels
-        unet.register_to_config(in_channels=in_channels)
-
-        with torch.no_grad():
-            new_conv_in = nn.Conv2d(
-                in_channels, out_channels, unet.conv_in.kernel_size, unet.conv_in.stride, unet.conv_in.padding
-            )
-            new_conv_in.weight.zero_()
-            new_conv_in.weight[:, :4, :, :].copy_(unet.conv_in.weight)
-            unet.conv_in = new_conv_in
+    logger.info("Initializing the InstructPix2Pix UNet from the pretrained UNet.")
+    in_channels = 8
+    out_channels = unet.conv_in.out_channels
+    unet.register_to_config(in_channels=in_channels)
+
+    with torch.no_grad():
+        new_conv_in = nn.Conv2d(
+            in_channels, out_channels, unet.conv_in.kernel_size, unet.conv_in.stride, unet.conv_in.padding
+        )
+        new_conv_in.weight.zero_()
+        new_conv_in.weight[:, :4, :, :].copy_(unet.conv_in.weight)
+        unet.conv_in = new_conv_in
 
     # Freeze vae and text_encoder
     vae.requires_grad_(False)
@@ -892,9 +891,12 @@ def collate_fn(examples):
                     # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
                     ema_unet.store(unet.parameters())
                     ema_unet.copy_to(unet.parameters())
+                # The models need unwrapping because for compatibility in distributed training mode.
                 pipeline = StableDiffusionInstructPix2PixPipeline.from_pretrained(
                     args.pretrained_model_name_or_path,
-                    unet=unet,
+                    unet=accelerator.unwrap_model(unet),
+                    text_encoder=accelerator.unwrap_model(text_encoder),
+                    vae=accelerator.unwrap_model(vae),
                     revision=args.revision,
                     torch_dtype=weight_dtype,
                 )
@@ -904,7 +906,9 @@ def collate_fn(examples):
                 # run inference
                 original_image = download_image(args.val_image_url)
                 edited_images = []
-                with torch.autocast(str(accelerator.device), enabled=accelerator.mixed_precision == "fp16"):
+                with torch.autocast(
+                    str(accelerator.device).replace(":0", ""), enabled=accelerator.mixed_precision == "fp16"
+                ):
                     for _ in range(args.num_validation_images):
                         edited_images.append(
                             pipeline(
@@ -959,7 +963,7 @@ def collate_fn(examples):
         if args.validation_prompt is not None:
             edited_images = []
             pipeline = pipeline.to(accelerator.device)
-            with torch.autocast(str(accelerator.device)):
+            with torch.autocast(str(accelerator.device).replace(":0", "")):
                 for _ in range(args.num_validation_images):
                     edited_images.append(
                         pipeline(