diffusers/docs/source/en/optimization/fp16.mdx

<!--Copyright 2022 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.
-->

# Memory and speed

We present some techniques and ideas to optimize 🤗 Diffusers _inference_ for memory or speed. As a general rule, we recommend the use of [xFormers](https://github.com/facebookresearch/xformers) for memory efficient attention, please see the recommended [installation instructions](xformers).

We'll discuss how the following settings impact performance and memory.

|                  | Latency | Speedup |
| ---------------- | ------- | ------- |
| original         | 9.50s   | x1      |
| cuDNN auto-tuner | 9.37s   | x1.01   |
| fp16             | 3.61s   | x2.63   |
| channels last    | 3.30s   | x2.88   |
| traced UNet      | 3.21s   | x2.96   |
| memory efficient attention  | 2.63s  | x3.61   |

<em>
  obtained on NVIDIA TITAN RTX by generating a single image of size 512x512 from
  the prompt "a photo of an astronaut riding a horse on mars" with 50 DDIM
  steps.
</em>

## Enable cuDNN auto-tuner

[NVIDIA cuDNN](https://developer.nvidia.com/cudnn) supports many algorithms to compute a convolution. Autotuner runs a short benchmark and selects the kernel with the best performance on a given hardware for a given input size.

Since we’re using **convolutional networks** (other types currently not supported), we can enable cuDNN autotuner before launching the inference by setting:

```python
import torch

torch.backends.cudnn.benchmark = True
```

### Use tf32 instead of fp32 (on Ampere and later CUDA devices)

On Ampere and later CUDA devices matrix multiplications and convolutions can use the TensorFloat32 (TF32) mode for faster but slightly less accurate computations. By default PyTorch enables TF32 mode for convolutions but not matrix multiplications, and unless a network requires full float32 precision we recommend enabling this setting for matrix multiplications, too. It can significantly speed up computations with typically negligible loss of numerical accuracy. You can read more about it [here](https://huggingface.co/docs/transformers/v4.18.0/en/performance#tf32). All you need to do is to add this before your inference:

```python
import torch

torch.backends.cuda.matmul.allow_tf32 = True
```

## Half precision weights

To save more GPU memory and get more speed, you can load and run the model weights directly in half precision. This involves loading the float16 version of the weights, which was saved to a branch named `fp16`, and telling PyTorch to use the `float16` type when loading them:

```Python
pipe = StableDiffusionPipeline.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]
```

<Tip warning={true}>
  It is strongly discouraged to make use of [`torch.autocast`](https://pytorch.org/docs/stable/amp.html#torch.autocast) in any of the pipelines as it can lead to black images and is always slower than using pure 
  float16 precision.
</Tip>

## Sliced attention for additional memory savings

For even additional memory savings, you can use a sliced version of attention that performs the computation in steps instead of all at once.

<Tip>
  Attention slicing is useful even if a batch size of just 1 is used - as long
  as the model uses more than one attention head. If there is more than one
  attention head the *QK^T* attention matrix can be computed sequentially for
  each head which can save a significant amount of memory.
</Tip>

To perform the attention computation sequentially over each head, you only need to invoke [`~StableDiffusionPipeline.enable_attention_slicing`] in your pipeline before inference, like here:

```Python
import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.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"
pipe.enable_attention_slicing()
image = pipe(prompt).images[0]
```

There's a small performance penalty of about 10% slower inference times, but this method allows you to use Stable Diffusion in as little as 3.2 GB of VRAM!


## Sliced VAE decode for larger batches

To decode large batches of images with limited VRAM, or to enable batches with 32 images or more, you can use sliced VAE decode that decodes the batch latents one image at a time.

You likely want to couple this with [`~StableDiffusionPipeline.enable_attention_slicing`] or [`~StableDiffusionPipeline.enable_xformers_memory_efficient_attention`] to further minimize memory use.

To perform the VAE decode one image at a time, invoke [`~StableDiffusionPipeline.enable_vae_slicing`] in your pipeline before inference. For example:

```Python
import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.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"
pipe.enable_vae_slicing()
images = pipe([prompt] * 32).images
```

You may see a small performance boost in VAE decode on multi-image batches. There should be no performance impact on single-image batches.


## Offloading to CPU with accelerate for memory savings

For additional memory savings, you can offload the weights to CPU and only load them to GPU when performing the forward pass.

To perform CPU offloading, all you have to do is invoke [`~StableDiffusionPipeline.enable_sequential_cpu_offload`]:

```Python
import torch
from diffusers import StableDiffusionPipeline

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

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

And you can get the memory consumption to < 3GB.

If is also possible to chain it with attention slicing for minimal memory consumption (< 2GB).

```Python
import torch
from diffusers import StableDiffusionPipeline

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

prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_sequential_cpu_offload()
pipe.enable_attention_slicing(1)

image = pipe(prompt).images[0]
```

**Note**: When using `enable_sequential_cpu_offload()`, it is important to **not** move the pipeline to CUDA beforehand or else the gain in memory consumption will only be minimal. See [this issue](https://github.com/huggingface/diffusers/issues/1934) for more information.

## Using Channels Last memory format

Channels last memory format is an alternative way of ordering NCHW tensors in memory preserving dimensions ordering. Channels last tensors ordered in such a way that channels become the densest dimension (aka storing images pixel-per-pixel). Since not all operators currently support channels last format it may result in a worst performance, so it's better to try it and see if it works for your model.

For example, in order to set the UNet model in our pipeline to use channels last format, we can use the following:

```python
print(pipe.unet.conv_out.state_dict()["weight"].stride())  # (2880, 9, 3, 1)
pipe.unet.to(memory_format=torch.channels_last)  # in-place operation
print(
    pipe.unet.conv_out.state_dict()["weight"].stride()
)  # (2880, 1, 960, 320) having a stride of 1 for the 2nd dimension proves that it works
```

## Tracing

Tracing runs an example input tensor through your model, and captures the operations that are invoked as that input makes its way through the model's layers so that an executable or `ScriptFunction` is returned that will be optimized using just-in-time compilation.

To trace our UNet model, we can use the following:

```python
import time
import torch
from diffusers import StableDiffusionPipeline
import functools

# torch disable grad
torch.set_grad_enabled(False)

# set variables
n_experiments = 2
unet_runs_per_experiment = 50


# load inputs
def generate_inputs():
    sample = torch.randn(2, 4, 64, 64).half().cuda()
    timestep = torch.rand(1).half().cuda() * 999
    encoder_hidden_states = torch.randn(2, 77, 768).half().cuda()
    return sample, timestep, encoder_hidden_states


pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
).to("cuda")
unet = pipe.unet
unet.eval()
unet.to(memory_format=torch.channels_last)  # use channels_last memory format
unet.forward = functools.partial(unet.forward, return_dict=False)  # set return_dict=False as default

# warmup
for _ in range(3):
    with torch.inference_mode():
        inputs = generate_inputs()
        orig_output = unet(*inputs)

# trace
print("tracing..")
unet_traced = torch.jit.trace(unet, inputs)
unet_traced.eval()
print("done tracing")


# warmup and optimize graph
for _ in range(5):
    with torch.inference_mode():
        inputs = generate_inputs()
        orig_output = unet_traced(*inputs)


# benchmarking
with torch.inference_mode():
    for _ in range(n_experiments):
        torch.cuda.synchronize()
        start_time = time.time()
        for _ in range(unet_runs_per_experiment):
            orig_output = unet_traced(*inputs)
        torch.cuda.synchronize()
        print(f"unet traced inference took {time.time() - start_time:.2f} seconds")
    for _ in range(n_experiments):
        torch.cuda.synchronize()
        start_time = time.time()
        for _ in range(unet_runs_per_experiment):
            orig_output = unet(*inputs)
        torch.cuda.synchronize()
        print(f"unet inference took {time.time() - start_time:.2f} seconds")

# save the model
unet_traced.save("unet_traced.pt")
```

Then we can replace the `unet` attribute of the pipeline with the traced model like the following

```python
from diffusers import StableDiffusionPipeline
import torch
from dataclasses import dataclass


@dataclass
class UNet2DConditionOutput:
    sample: torch.FloatTensor


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

# use jitted unet
unet_traced = torch.jit.load("unet_traced.pt")


# del pipe.unet
class TracedUNet(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.in_channels = pipe.unet.in_channels
        self.device = pipe.unet.device

    def forward(self, latent_model_input, t, encoder_hidden_states):
        sample = unet_traced(latent_model_input, t, encoder_hidden_states)[0]
        return UNet2DConditionOutput(sample=sample)


pipe.unet = TracedUNet()

with torch.inference_mode():
    image = pipe([prompt] * 1, num_inference_steps=50).images[0]
```


## Memory Efficient Attention

Recent work on optimizing the bandwitdh in the attention block has generated huge speed ups and gains in GPU memory usage. The most recent being Flash Attention from @tridao: [code](https://github.com/HazyResearch/flash-attention), [paper](https://arxiv.org/pdf/2205.14135.pdf).

Here are the speedups we obtain on a few Nvidia GPUs when running the inference at 512x512 with a batch size of 1 (one prompt):

| GPU              	| Base Attention FP16 	| Memory Efficient Attention FP16 	|
|------------------	|---------------------	|---------------------------------	|
| NVIDIA Tesla T4  	| 3.5it/s             	| 5.5it/s                         	|
| NVIDIA 3060 RTX  	| 4.6it/s             	| 7.8it/s                         	|
| NVIDIA A10G      	| 8.88it/s            	| 15.6it/s                        	|
| NVIDIA RTX A6000 	| 11.7it/s            	| 21.09it/s                       	|
| NVIDIA TITAN RTX  | 12.51it/s         	| 18.22it/s                       	|
| A100-SXM4-40GB    	| 18.6it/s            	| 29.it/s                        	|
| A100-SXM-80GB    	| 18.7it/s            	| 29.5it/s                        	|

To leverage it just make sure you have: 
 - PyTorch > 1.12
 - Cuda available
 - [Installed the xformers library](xformers).
```python
from diffusers import StableDiffusionPipeline
import torch

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

pipe.enable_xformers_memory_efficient_attention()

with torch.inference_mode():
    sample = pipe("a small cat")

# optional: You can disable it via
# pipe.disable_xformers_memory_efficient_attention()
```