Document sequential CPU offload method on Stable Diffusion pipeline (#1024)

* document cpu offloading method

* address review comments
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

docs/source/optimization/fp16.mdx

@@ -14,9 +14,8 @@ specific language governing permissions and limitations under the License.
 
 We present some techniques and ideas to optimize 🤗 Diffusers _inference_ for memory or speed.
 
-
 |                  | Latency | Speedup |
-|------------------|---------|---------|
+| ---------------- | ------- | ------- |
 | original         | 9.50s   | x1      |
 | cuDNN auto-tuner | 9.37s   | x1.01   |
 | autocast (fp16)  | 5.47s   | x1.91   |
@@ -24,7 +23,11 @@ We present some techniques and ideas to optimize 🤗 Diffusers _inference_ for
 | channels last    | 3.30s   | x2.87   |
 | traced UNet      | 3.21s   | x2.96   |
 
-<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>
+<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
 
@@ -87,7 +90,10 @@ image = pipe(prompt).images[0]
 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.
+  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:
@@ -110,6 +116,50 @@ 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!
 
+## Offloading to CPU with accelerate for memory savings
+
+For additional memory savings, you can offload the weights to CPU and 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",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+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 < 2GB.
+
+If is also possible to chain it with attention slicing for minimal memory consumption, running it in as little as < 800mb of GPU vRAM:
+
+```Python
+import torch
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    revision="fp16",
+    torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+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]
+```
+
 ## 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.

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -120,6 +120,11 @@ class StableDiffusionPipeline(DiffusionPipeline):
         self.enable_attention_slicing(None)
 
     def enable_sequential_cpu_offload(self):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+        text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+        `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+        """
         if is_accelerate_available():
             from accelerate import cpu_offload
         else: