[Draft] GPU Support Document

[iwr-redmond]

GPU Support

While dgenerate supports most graphics cards, NVIDIA GPUs are recommended for the highest level of compatibility. At least 8GB of video memory (VRAM) is suggested.

GPU support summary

GPU Type	Current Support	Torch Device
NVIDIA (CUDA)	Full	cuda
Intel Arc (XPU)	Partial*	xpu
AMD (ROCm)	Partial	cuda
Apple Silicon (MPS)	Partial*	mps
CPU	Limited	cpu

Notes:

• Intel Arc (XPU) and Apple Silicon (MPS) are both experimentally supported in version 5.0.0; please consider filing a bug report if you encounter any issues
• Using the CPU is extremely slow for all generation tasks; selecting the LCM Scheduler, Hyper-SD LoRAs (SD-1step, SDXL-1step, SD3-4step, Flux-4step), and Tiny VAEs (SD, SDXL, SD3, Flux) will provide the best performance on systems that lack a dedicated GPU

Rendering compatibility by GPU type

Card Type	Windows	Linux	macOS
NVIDIA (CUDA)	✅	✅	❌
Intel Arc (XPU)	✅	✅	❌
AMD (ROCm)	❌	✅	❌
Apple Silicon (MPS)	❌	❌	✅
CPU (limited support)	✅	✅	✅

Notes:

• Support for AMD (ROCm) on Windows is planned for a future release (see upstream status)
• Support for legacy Apple Intel and Windows DirectML architectures is not planned

Quantization compatibility by GPU type

Card Type	Bitsandbytes	SDNQ	GGUF
NVIDIA (CUDA)	✅	✅	❌
Intel Arc (XPU)	❌	✅	❌
AMD (ROCm)	❌	✅	❌
Apple Silicon (MPS)	❌	✅	❌
CPU (limited support)	✅*	✅	❌

Notes:

• Support for Bitsandbytes on Intel Arc (XPU), AMD (ROCm), and Apple Silicon (MPS) is planned for a future release (see upstream status)
• Bitsandbytes is not yet supported when using the CPU on Apple systems
• Support for GGUF diffusion checkpoints is not planned (read why); however, GGUF support for some LLMs is available when using the GPT4All prompt upscaler