<!--Copyright 2023 The HuggingFace Team. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->

# Accelerated PyTorch 2.0 support in Diffusers

Starting from version `0.13.0`, Diffusers supports the latest optimization from the upcoming [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/) release. These include:
1. Support for accelerated transformers implementation with memory-efficient attention – no extra dependencies required.
2. [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) support for extra performance boost when individual models are compiled.


## Installation
To benefit from the accelerated attention implementation and `torch.compile`, you just need to install the latest versions of PyTorch 2.0 from `pip`, and make sure you are on diffusers 0.13.0 or later. As explained below, `diffusers` automatically uses the attention optimizations (but not `torch.compile`) when available.

```bash
pip install --upgrade torch torchvision diffusers
```

## Using accelerated transformers and torch.compile.


1. **Accelerated Transformers implementation**

   PyTorch 2.0 includes an optimized and memory-efficient attention implementation through the [`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) function, which automatically enables several optimizations depending on the inputs and the GPU type. This is similar to the `memory_efficient_attention` from [xFormers](https://github.com/facebookresearch/xformers), but built natively into PyTorch. 

   These optimizations will be enabled by default in Diffusers if PyTorch 2.0 is installed and if `torch.nn.functional.scaled_dot_product_attention` is available. To use it, just install `torch 2.0` as suggested above and simply use the pipeline. For example:

    ```Python
    import torch
    from diffusers import DiffusionPipeline

    pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
    pipe = pipe.to("cuda")

    prompt = "a photo of an astronaut riding a horse on mars"
    image = pipe(prompt).images[0]
    ```

    If you want to enable it explicitly (which is not required), you can do so as shown below.

    ```Python
    import torch
    from diffusers import DiffusionPipeline
    from diffusers.models.attention_processor import AttnProcessor2_0

    pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
    pipe.unet.set_attn_processor(AttnProcessor2_0())

    prompt = "a photo of an astronaut riding a horse on mars"
    image = pipe(prompt).images[0]
    ```

    This should be as fast and memory efficient as `xFormers`. More details [in our benchmark](#benchmark).


2. **torch.compile**

    To get an additional speedup, we can use the new `torch.compile` feature. To do so, we simply wrap our `unet` with `torch.compile`. For more information and different options, refer to the 
    [torch compile docs](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html).

    ```python
    import torch
    from diffusers import DiffusionPipeline

    pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
    pipe.unet = torch.compile(pipe.unet)

    batch_size = 10
    prompt = "A photo of an astronaut riding a horse on marse."
    images = pipe(prompt, num_inference_steps=steps, num_images_per_prompt=batch_size).images
    ```

    Depending on the type of GPU, `compile()` can yield between 2-9% of _additional speed-up_ over the accelerated transformer optimizations. Note, however, that compilation is able to squeeze more performance improvements in more recent GPU architectures such as Ampere (A100, 3090), Ada (4090) and Hopper (H100).
    
    Compilation takes some time to complete, so it is best suited for situations where you need to prepare your pipeline once and then perform the same type of inference operations multiple times.


## Benchmark

We conducted a simple benchmark on different GPUs to compare vanilla attention, xFormers, `torch.nn.functional.scaled_dot_product_attention` and `torch.compile+torch.nn.functional.scaled_dot_product_attention`.
For the benchmark we used the [stable-diffusion-v1-4](https://huggingface.co/CompVis/stable-diffusion-v1-4) model with 50 steps. The `xFormers` benchmark is done using the `torch==1.13.1` version, while the accelerated transformers optimizations are tested using nightly versions of PyTorch 2.0. The tables below summarize the results we got.

Please refer to [our featured blog post in the PyTorch site](https://pytorch.org/blog/accelerated-diffusers-pt-20/) for more details.

### FP16 benchmark

The table below shows the benchmark results for inference using `fp16`. As we can see, `torch.nn.functional.scaled_dot_product_attention` is as fast as `xFormers` (sometimes slightly faster/slower) on all the GPUs we tested.
And using `torch.compile` gives further speed-up of up of 10% over `xFormers`, but it's mostly noticeable on the A100 GPU.

___The time reported is in seconds.___

| GPU | Batch Size | Vanilla Attention | xFormers | PyTorch2.0 SDPA | SDPA + torch.compile | Speed over xformers (%) |
| --- | --- | --- | --- | --- | --- | --- |
| A100 | 1 | 2.69 | 2.7 | 1.98 | 2.47 | 8.52 |
| A100 | 2 | 3.21 | 3.04 | 2.38 | 2.78 | 8.55 |
| A100 | 4 | 5.27 | 3.91 | 3.89 | 3.53 | 9.72 |
| A100 | 8 | 9.74 | 7.03 | 7.04 | 6.62 | 5.83 |
| A100 | 10 | 12.02 | 8.7 | 8.67 | 8.45 | 2.87 |
| A100 | 16 | 18.95 | 13.57 | 13.55 | 13.20 | 2.73 |
| A100 | 32 (1) | OOM | 26.56 | 26.68 | 25.85 | 2.67 |
| A100 | 64 | | 52.51 | 53.03 | 50.93 | 3.01 |
| | | | | | | |
| A10 | 4 | 13.94 | 9.81 | 10.01 | 9.35 | 4.69 |
| A10 | 8 | 27.09 | 19 | 19.53 | 18.33 | 3.53 |
| A10 | 10 | 33.69 | 23.53 | 24.19 | 22.52 | 4.29 |
| A10 | 16 | OOM | 37.55 | 38.31 | 36.81 | 1.97 |
| A10 | 32 (1) | | 77.19 | 78.43 | 76.64 | 0.71 |
| A10 | 64 (1) | | 173.59 | 158.99 | 155.14 | 10.63 |
| | | | | | | |
| T4 | 4 | 38.81 | 30.09 | 29.74 | 27.55 | 8.44 |
| T4 | 8 | OOM | 55.71 | 55.99 | 53.85 | 3.34 |
| T4 | 10 | OOM | 68.96 | 69.86 | 65.35 | 5.23 |
| T4 | 16 | OOM | 111.47 | 113.26  | 106.93  | 4.07 |
| | | | | | | |
| V100 | 4 | 9.84 | 8.16 | 8.09 | 7.65 | 6.25 |
| V100 | 8 | OOM | 15.62 | 15.44 | 14.59 | 6.59 |
| V100 | 10 | OOM | 19.52 | 19.28 | 18.18 | 6.86 |
| V100 | 16 | OOM | 30.29 | 29.84 | 28.22 | 6.83 |
| | | | | | | |
| 3090 | 1 | 2.94 | 2.5 | 2.42 | 2.33 | 6.80 |
| 3090 | 4 | 10.04 | 7.82 | 7.72 | 7.38 | 5.63 |
| 3090 | 8 | 19.27 | 14.97 | 14.88 | 14.15 | 5.48 |
| 3090 | 10| 24.08 | 18.7 | 18.62 | 18.12 | 3.10 |
| 3090 | 16 | OOM | 29.06 | 28.88 | 28.2 | 2.96 |
| 3090 | 32 (1) | | 58.05 | 57.42 | 56.28 | 3.05 |
| 3090 | 64 (1) | | 126.54 | 114.27 | 112.21 | 11.32 |
| | | | | | | |
| 3090 Ti | 1 | 2.7 | 2.26 | 2.19 | 2.12 | 6.19 |
| 3090 Ti | 4 | 9.07 | 7.14 | 7.00 | 6.71 | 6.02 |
| 3090 Ti | 8 | 17.51 | 13.65 | 13.53 | 12.94 | 5.20 |
| 3090 Ti | 10 (2) | 21.79 | 16.85 | 16.77 | 16.44 | 2.43 |
| 3090 Ti | 16 | OOM | 26.1 | 26.04 | 25.53 | 2.18 |
| 3090 Ti | 32 (1) | | 51.78 | 51.71 | 50.91 | 1.68 |
| 3090 Ti | 64 (1) | | 112.02 | 102.78 | 100.89 | 9.94 |
| | | | | | | |
| 4090 | 1 | 4.47 | 3.98 | 1.28 | 1.21 | 69.60 |
| 4090 | 4 | 10.48 | 8.37 | 3.76 | 3.56 | 57.47 |
| 4090 | 8 | 14.33 | 10.22 | 7.43 | 6.99 | 31.60 |
| 4090 | 16 | | 17.07 | 14.98 | 14.58 | 14.59 |
| 4090 | 32 (1) | | 39.03 | 30.18 | 29.49 | 24.44 |
| 4090 | 64 (1) | | 77.29 | 61.34 | 59.96 | 22.42 |


### FP32 benchmark

The table below shows the benchmark results for inference using `fp32`. In this case, `torch.nn.functional.scaled_dot_product_attention` is faster than `xFormers` on all the GPUs we tested.

Using `torch.compile` in addition to the accelerated transformers implementation can yield up to 19% performance improvement over `xFormers` in Ampere and Ada cards, and up to 20% (Ampere) or 28% (Ada) over vanilla attention.

| GPU | Batch Size | Vanilla Attention | xFormers | PyTorch2.0 SDPA | SDPA + torch.compile | Speed over xformers (%) | Speed over vanilla (%) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| A100 | 1 | 4.97 | 3.86 | 2.6 | 2.86 | 25.91 | 42.45 |
| A100 | 2 | 9.03 | 6.76 | 4.41 | 4.21 | 37.72 | 53.38 |
| A100 | 4 | 16.70 | 12.42 | 7.94 | 7.54 | 39.29 | 54.85 |
| A100 | 10 | OOM | 29.93 | 18.70 | 18.46 | 38.32 | |
| A100 | 16 | | 47.08 | 29.41 | 29.04 | 38.32 | |
| A100 | 32 | | 92.89 | 57.55 | 56.67 | 38.99 | |
| A100 | 64 | | 185.3 | 114.8 | 112.98 | 39.03 | |
| | | | | | | |
| A10 | 1 | 10.59 | 8.81 | 7.51 | 7.35 | 16.57 | 30.59 |
| A10 | 4 | 34.77 | 27.63 | 22.77 | 22.07 | 20.12 | 36.53 |
| A10 | 8 | | 56.19 | 43.53 | 43.86 | 21.94 | |
| A10 | 16 | | 116.49 | 88.56 | 86.64 | 25.62 | |
| A10 | 32 | | 221.95 | 175.74 | 168.18 | 24.23 | |
| A10 | 48 | | 333.23 | 264.84 | | 20.52 | |
| | | | | | | |
| T4 | 1 | 28.2 | 24.49 | 23.93 | 23.56 | 3.80 | 16.45 |
| T4 | 2 | 52.77 | 45.7 | 45.88 | 45.06 | 1.40 | 14.61 |
| T4 | 4 | OOM | 85.72 | 85.78 | 84.48 | 1.45 | |
| T4 | 8 | | 149.64 | 150.75 | 148.4 | 0.83 | |
| | | | | | | |
| V100 | 1 | 7.4 | 6.84 | 6.8 | 6.66 | 2.63 | 10.00 |
| V100 | 2 | 13.85 | 12.81 | 12.66 | 12.35 | 3.59 | 10.83 |
| V100 | 4 | OOM | 25.73 | 25.31 | 24.78 | 3.69 | |
| V100 | 8 | | 43.95 | 43.37 | 42.25 | 3.87 | |
| V100 | 16 | | 84.99 | 84.73 | 82.55 | 2.87 | |
| | | | | | | |
| 3090 | 1 | 7.09 | 6.78 | 5.34 | 5.35 | 21.09 | 24.54 |
| 3090 | 4 | 22.69 | 21.45 | 18.56 | 18.18 | 15.24 | 19.88 |
| 3090 | 8 | | 42.59 | 36.68 | 35.61 | 16.39 | |
| 3090 | 16 | | 85.35 | 72.93 | 70.18 | 17.77 | |
| 3090 | 32 (1) | | 162.05 | 143.46 | 138.67 | 14.43 | |
| | | | | | | |
| 3090 Ti | 1 | 6.45 | 6.19 | 4.99 | 4.89 | 21.00 | 24.19 |
| 3090 Ti | 4 | 20.32 | 19.31 | 17.02 | 16.48 | 14.66 | 18.90 |
| 3090 Ti | 8 | | 37.93 | 33.21 | 32.24 | 15.00 | |
| 3090 Ti | 16 | | 75.37 | 66.63 | 64.5 | 14.42 | |
| 3090 Ti | 32 (1) | | 142.55 | 128.89 | 124.92 | 12.37 | |
| | | | | | | |
| 4090 | 1 | 5.54 | 4.99 | 2.66 | 2.58 | 48.30 | 53.43 |
| 4090 | 4 | 13.67 | 11.4 | 8.81 | 8.46 | 25.79 | 38.11 |
| 4090 | 8 |  | 19.79 | 17.55 | 16.62 | 16.02 |  |
| 4090 | 16 |  | 38.62 | 35.65 | 34.07 | 11.78 |  |
| 4090 | 32 (1) |  | 76.57 | 69.48 | 65.35 | 14.65 |  |
| 4090 | 48 |  | 114.44 | 106.3 |  | 7.11 |  |


(1) Batch Size >= 32 requires enable_vae_slicing() because of https://github.com/pytorch/pytorch/issues/81665.
This is required for PyTorch 1.13.1, and also for PyTorch 2.0 and large batch sizes.

For more details about how this benchmark was run, please refer to [this PR](https://github.com/huggingface/diffusers/pull/2303) and to [the blog post](https://pytorch.org/blog/accelerated-diffusers-pt-20/).