perf: accelerate INT8 diffusion inference with CUDA fusions (#2235)

docs/docker.md

@@ -18,12 +18,26 @@ docker run --rm --init -v /path/to/models:/models -v /path/to/output/:/output -p
 
 ## Building using Docker
 
+Initialize the submodules from the repository root before building. Docker copies
+the local checkout, so its build context must include the submodule contents:
+
 ```shell
+git submodule update --init --recursive
 docker build -f docker/Dockerfile -t sd .
 ```
 
 ## Building variants using Docker
 
+### CUDA (NVIDIA GPU):
+
+```shell
+docker build -f docker/Dockerfile.cuda -t sd-cuda .
+```
+
+To select GPU architectures, pass `--build-arg "CUDA_ARCHITECTURES=89"` (for
+example, RTX 4090). Run the image on a host configured with NVIDIA Container
+Toolkit, using `docker run --rm --gpus all ... sd-cuda [args...]`.
+
 ### Vulkan:
 
 ```shell
@@ -67,4 +81,4 @@ docker run --rm --device=/dev/dri --device=/dev/kfd --group-add video \
 #   -v ./models:/models -v ./build:/output \
 #   sd -m /models/sd-v1-4.ckpt -v -o /output/output.png \
 #   -p "a lovely cat"
-```
+```