104 lines
3 KiB
Text
104 lines
3 KiB
Text
Usage
|
|
-----
|
|
|
|
1. Check the Linux Kernel Module (driver) for Intel GPU:
|
|
|
|
lsmod | grep -i xe
|
|
|
|
For desktop distros with rolling-ish kernel updates, xe is usually included already.
|
|
|
|
2. Search and install `intel-compute-runtime` (or similar) with your host OS' package manager.
|
|
If not found, just ignore it and run directly.
|
|
|
|
For Ubuntu, follow this instruction and install the compute-related packages:
|
|
https://dgpu-docs.intel.com/driver/client/overview.html
|
|
|
|
3. Run the container:
|
|
|
|
====
|
|
mkdir -p \
|
|
storage \
|
|
storage-models/models \
|
|
storage-models/hf-hub \
|
|
storage-models/torch-hub \
|
|
storage-user/input \
|
|
storage-user/output \
|
|
storage-user/workflows
|
|
|
|
podman run -it --rm \
|
|
--name comfyui-xpu \
|
|
--device=/dev/dri \
|
|
--ipc=host \
|
|
--security-opt label=disable \
|
|
-p 8188:8188 \
|
|
-v "$(pwd)"/storage:/root \
|
|
-v "$(pwd)"/storage-models/models:/root/ComfyUI/models \
|
|
-v "$(pwd)"/storage-models/hf-hub:/root/.cache/huggingface/hub \
|
|
-v "$(pwd)"/storage-models/torch-hub:/root/.cache/torch/hub \
|
|
-v "$(pwd)"/storage-user/input:/root/ComfyUI/input \
|
|
-v "$(pwd)"/storage-user/output:/root/ComfyUI/output \
|
|
-v "$(pwd)"/storage-user/workflows:/root/ComfyUI/user/default/workflows \
|
|
-e CLI_ARGS="--disable-smart-memory --async-offload" \
|
|
yanwk/comfyui-boot:xpu
|
|
====
|
|
|
|
Note:
|
|
|
|
1. "--disable-smart-memory" is useful for XPU, although it consumes more RAM.
|
|
2. "--async-offload" gives subtle improvement and no visible harm.
|
|
|
|
More CLI_ARGS:
|
|
|
|
--lowvram
|
|
--bf16-unet
|
|
--bf16-vae
|
|
--bf16-text-enc
|
|
--mmap-torch-files
|
|
--reserve-vram 1
|
|
|
|
Check the doc before use:
|
|
https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cli_args.py
|
|
|
|
|
|
Test result on Arc B580
|
|
-----------------------
|
|
|
|
Time on generating one 1024x1024px image (pre-warmed, only infer time)
|
|
|
|
* SD 1.5 (20-step): OK (5s) | 512x512: GOOD (1s)
|
|
* SDXL (28-step): OK (6s)
|
|
* SDXL Refiner (20+5-step): OK (4+1s)
|
|
* Stable Cascade (20+10-step): FAST (3+3s)
|
|
* SD 3.5 large fp8 (20-step): GOOD (26s)
|
|
* Flux1 schnell (4-step): GOOD (8s)
|
|
* Flux1 dev fp8 (20-step): GOOD (35s)
|
|
* Flux1 Krea dev (20-step): OK (46s)
|
|
|
|
* Hunyuan3D 2.0 (Comfy Repackaged): 'Non-uniform work-groups are not supported by the target device'
|
|
* StableZero123: 'Non-uniform work-groups are not supported by the target device'
|
|
|
|
* Wan 2.2 5B Text to Video (640x352, 121 frames): FAST (60s, very poor quality)
|
|
* Wan 2.2 5B Text to Video (960x544, 121 frames): OK (181s, low quality)
|
|
* Wan 2.2 5B Text to Video (1280x704, 121 frames): FAILED (OOM)
|
|
|
|
|
|
Notes
|
|
--------------
|
|
|
|
1. Once OOM, ComfyUI won't crash, but GPU will stop responding
|
|
for future prompt (UR_RESULT_ERROR_DEVICE_LOST).
|
|
In that case, you will need to restart ComfyUI.
|
|
|
|
2. IPEX (Intel Extension for PyTorch) is not included, as most of its features
|
|
have been merged upstream and it is being phased out:
|
|
https://github.com/intel/intel-extension-for-pytorch/releases
|
|
|
|
|
|
For Windows Users
|
|
-----------------
|
|
|
|
Just use Intel AI Playground:
|
|
https://game.intel.com/us/stories/introducing-ai-playground/
|
|
|
|
Or yet another ComfyUI portable:
|
|
https://github.com/YanWenKun/ComfyUI-WinPortable-XPU
|