Make xformers optional even if it is available (#1753)

* Make xformers optional even if it is available

* Raise exception if xformers is used but not available

* Rename use_xformers to enable_xformers_memory_efficient_attention

* Add a note about xformers in README

* Reformat code style

examples/dreambooth/README.md

@@ -317,4 +317,7 @@ python train_dreambooth_flax.py \
   --max_train_steps=800
 ```
 
-You can also use Dreambooth to train the specialized in-painting model. See [the script in the research folder for details](https://github.com/huggingface/diffusers/tree/main/examples/research_projects/dreambooth_inpaint).
+### Training with xformers:
\ No newline at end of file
+You can enable memory efficient attention by [installing xFormers](https://github.com/facebookresearch/xformers#installing-xformers) and padding the `--enable_xformers_memory_efficient_attention` argument to the script. This is not available with the Flax/JAX implementation.
+
+You can also use Dreambooth to train the specialized in-painting model. See [the script in the research folder for details](https://github.com/huggingface/diffusers/tree/main/examples/research_projects/dreambooth_inpaint).

examples/dreambooth/train_dreambooth.py

@@ -248,6 +248,9 @@ def parse_args(input_args=None):
         ),
     )
     parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -516,14 +519,11 @@ def main(args):
         revision=args.revision,
     )
 
-    if is_xformers_available():
-        try:
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
             unet.enable_xformers_memory_efficient_attention()
-        except Exception as e:
-            logger.warning(
-                "Could not enable memory efficient attention. Make sure xformers is installed"
-                f" correctly and a GPU is available: {e}"
-            )
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
 
     vae.requires_grad_(False)
     if not args.train_text_encoder:

examples/text_to_image/README.md

@@ -160,3 +160,6 @@ python train_text_to_image_flax.py \
   --max_grad_norm=1 \
   --output_dir="sd-pokemon-model"
 ```
+
+### Training with xformers:
+You can enable memory efficient attention by [installing xFormers](https://github.com/facebookresearch/xformers#installing-xformers) and padding the `--enable_xformers_memory_efficient_attention` argument to the script. This is not available with the Flax/JAX implementation.

examples/text_to_image/train_text_to_image.py

@@ -234,6 +234,9 @@ def parse_args():
             ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
         ),
     )
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -383,14 +386,11 @@ def main():
         revision=args.revision,
     )
 
-    if is_xformers_available():
-        try:
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
             unet.enable_xformers_memory_efficient_attention()
-        except Exception as e:
-            logger.warning(
-                "Could not enable memory efficient attention. Make sure xformers is installed"
-                f" correctly and a GPU is available: {e}"
-            )
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
 
     # Freeze vae and text_encoder
     vae.requires_grad_(False)

examples/textual_inversion/README.md

@@ -124,3 +124,6 @@ python textual_inversion_flax.py \
   --output_dir="textual_inversion_cat"
 ```
 It should be at least 70% faster than the PyTorch script with the same configuration.
+
+### Training with xformers:
+You can enable memory efficient attention by [installing xFormers](https://github.com/facebookresearch/xformers#installing-xformers) and padding the `--enable_xformers_memory_efficient_attention` argument to the script. This is not available with the Flax/JAX implementation.

examples/textual_inversion/textual_inversion.py

@@ -222,6 +222,9 @@ def parse_args():
             ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
         ),
     )
+    parser.add_argument(
+        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
+    )
 
     args = parser.parse_args()
     env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
@@ -457,14 +460,11 @@ def main():
         revision=args.revision,
     )
 
-    if is_xformers_available():
-        try:
+    if args.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
             unet.enable_xformers_memory_efficient_attention()
-        except Exception as e:
-            logger.warning(
-                "Could not enable memory efficient attention. Make sure xformers is installed"
-                f" correctly and a GPU is available: {e}"
-            )
+        else:
+            raise ValueError("xformers is not available. Make sure it is installed correctly")
 
     # Resize the token embeddings as we are adding new special tokens to the tokenizer
     text_encoder.resize_token_embeddings(len(tokenizer))