[docs] PyTorch 2.0 (#11618)

* combine

* Update docs/source/en/optimization/fp16.md

Co-authored-by: Sayak Paul <[email protected]>

---------

Co-authored-by: Sayak Paul <[email protected]>

Author: stevhliu

docs/source/en/_toctree.yml

@@ -180,8 +180,6 @@
     title: Accelerate inference
   - local: optimization/memory
     title: Reduce memory usage
-  - local: optimization/torch2.0
-    title: PyTorch 2.0
   - local: optimization/xformers
     title: xFormers
   - local: optimization/tome

docs/source/en/api/pipelines/deepfloyd_if.md

@@ -347,7 +347,7 @@ pipe.to("cuda")
 image = pipe(image=image, prompt="<prompt>", strength=0.3).images
 ```
 
-You can also use [`torch.compile`](../../optimization/torch2.0). Note that we have not exhaustively tested `torch.compile`
+You can also use [`torch.compile`](../../optimization/fp16#torchcompile). Note that we have not exhaustively tested `torch.compile`
 with IF and it might not give expected results.
 
 ```py

docs/source/en/optimization/fp16.md

@@ -150,6 +150,24 @@ pipeline(prompt, num_inference_steps=30).images[0]
 
 Compilation is slow the first time, but once compiled, it is significantly faster. Try to only use the compiled pipeline on the same type of inference operations. Calling the compiled pipeline on a different image size retriggers compilation which is slow and inefficient.
 
+### Regional compilation
+
+[Regional compilation](https://docs.pytorch.org/tutorials/recipes/regional_compilation.html) reduces the cold start compilation time by only compiling a specific repeated region (or block) of the model instead of the entire model. The compiler reuses the cached and compiled code for the other blocks.
+
+[Accelerate](https://huggingface.co/docs/accelerate/index) provides the [compile_regions](https://github.com/huggingface/accelerate/blob/273799c85d849a1954a4f2e65767216eb37fa089/src/accelerate/utils/other.py#L78) method for automatically compiling the repeated blocks of a `nn.Module` sequentially. The rest of the model is compiled separately.
+
+```py
+# pip install -U accelerate
+import torch
+from diffusers import StableDiffusionXLPipeline
+from accelerate.utils import compile regions
+
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+).to("cuda")
+pipeline.unet = compile_regions(pipeline.unet, mode="reduce-overhead", fullgraph=True)
+```
+
 ### Graph breaks
 
 It is important to specify `fullgraph=True` in torch.compile to ensure there are no graph breaks in the underlying model. This allows you to take advantage of torch.compile without any performance degradation. For the UNet and VAE, this changes how you access the return variables.
@@ -170,6 +188,12 @@ The `step()` function is [called](https://github.com/huggingface/diffusers/blob/
 
 In general, the `sigmas` should [stay on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240) to avoid the communication sync and latency.
 
+### Benchmarks
+
+Refer to the [diffusers/benchmarks](https://huggingface.co/datasets/diffusers/benchmarks) dataset to see inference latency and memory usage data for compiled pipelines.
+
+The [diffusers-torchao](https://github.com/sayakpaul/diffusers-torchao#benchmarking-results) repository also contains benchmarking results for compiled versions of Flux and CogVideoX.
+
 ## Dynamic quantization
 
 [Dynamic quantization](https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html) improves inference speed by reducing precision to enable faster math operations. This particular type of quantization determines how to scale the activations based on the data at runtime rather than using a fixed scaling factor. As a result, the scaling factor is more accurately aligned with the data.

docs/source/en/optimization/tome.md

@@ -93,4 +93,4 @@ To reproduce this benchmark, feel free to use this [script](https://gist.github.
 |          |                |              2 |         OOM |             13 |               10.78 |
 |          |                |              1 |         OOM |           6.66 |                5.54 |
 
-As seen in the tables above, the speed-up from `tomesd` becomes more pronounced for larger image resolutions. It is also interesting to note that with `tomesd`, it is possible to run the pipeline on a higher resolution like 1024x1024. You may be able to speed-up inference even more with [`torch.compile`](torch2.0).
+As seen in the tables above, the speed-up from `tomesd` becomes more pronounced for larger image resolutions. It is also interesting to note that with `tomesd`, it is possible to run the pipeline on a higher resolution like 1024x1024. You may be able to speed-up inference even more with [`torch.compile`](fp16#torchcompile).