AMD support about fooocus

[kenwong1]

The minimum requirements that @lllyasviel posted above says 16GB GPU memory required for AMD on Windows with DirectML

[ahgera]

ComfyUI already has support for AMD cards on Windows via Microsoft's DirectML, but it has to be activated through a CLI argument (--directml). Since Fooocus imports Comfy directly I don't see any way of setting the args.directml value to anything from Fooocus' code.

I tried to hack around this in the Comfy repository. First by changing if args.directml is not None: to if not torch.cuda.is_available() and platform.system() == "Windows":, but Fooocus for some reason pulls down the latest Comfy version from git at every start, thereby overwriting any changes.

So I commented out that code, and after a while I got the web UI running, but i was unable to generate any images since I apparently have too little RAM in my old computer (12 GB):
RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 84934656 bytes.

The GPU and VRAM was barely used during this, so there was some bug still preventing DirectML to be used I think.

On top of that xFormers complains that it was built for PyTorch with Cuda, and I could find any way of either installing a CPU only version, or disable it.

---

I got slightly better results trying to use the same SDXL models in ComfyUI, but ultimately I got out of memory error there as well and gave up.
[W D:\a\_work\1\s\pytorch-directml-plugin\torch_directml\csrc\engine\dml_heap_allocator.cc:120] DML allocator out of memory!
At least this time i could see that both the GPU and VRAM was being used, but I'm guessing the limiting factor is still RAM since I got 8 GB VRAM and there were still some free.

I should mention that in both Fooocus' and ComfyUI's Python environments i had installed torch-directml (version 0.2.0.dev230426), which obviously is needed to import it.

[icecore2]

Same issue with RX 6700 XT.

[johan-lejdung]

Same on AMD 6750XT

[Small-Ku]

Related: comfyanonymous/ComfyUI#160

[chenshiwei-io]

Same on Amd RX588

[xiao-xiaozi]

same on Amd 580

[Roninos]

RX 5700XT not work

[StarLord-bot]

AMD Radeon RX 7900 XTX not working here as well

[SkyHype2000]

AMD Radeon RX 6750 XT also does not work here

[lllyasviel]

See also https://github.com/lllyasviel/Fooocus?tab=readme-ov-file#minimal-requirement

[kenwong1]

I've been using Fooocus on a 7900 XTX and it seems to work. Just need to make the changes mentioned in #624 (comment)

I would like to see some optimizations done to make it faster on AMD cards though

[SkyHype2000]

thanks it works now ^^

edited: RuntimeError: Could not allocate tensor with 167772160 bytes. There is not enough GPU video memory available!