multigpu_diffusion

Python Flask hosts for distributed parallel Diffusion inferencing solutions.

Model Compatibility Table:

Note: This table only indicates what is supported in this project - it is not indicative of what is supported in the software being implemented.

AsyncDiff:

Model	Type Name	LoRA	ControlNet	IPAdapter	Quantization	Compile	CPU Offloading
AnimateDiff	ad	❌	✅	✅	✅	✅❓	❌
Stable Diffusion 1.5	sd1	✅❓	✅❓	✅❓	✅❓	✅❓	❌
Stable Diffusion 2	sd2	✅❓	✅❓	✅❓	✅❓	✅❓	❌
Stable Diffusion 3	sd3	✅❓	✅❓	✅❓	✅❓	✅❓	❌
Stable Diffusion Upscaler	sdup	❌	❌	❌	✅❓	✅❓	❌
Stable Diffusion XL	sdxl	✅	✅	✅	✅	✅❓	❌
Stable Video Diffusion	svd	❌	❌	❌	✅❓	✅❓	❌

xDiT:

Model	Type Name	LoRA	ControlNet	IPAdapter	Quantization	Compile	CPU Offloading
FLUX.1	flux	✅❓	❌	✅	✅	✅	✅❓
HunYuan	hy	✅❓	❌	✅❓	✅❓	✅❓	✅❓
PixArt Alpha	pixa	✅❓	❌	✅❓	✅❓	✅❓	✅❓
PixArt Sigma	pixs	✅❓	❌	✅❓	✅❓	✅❓	✅❓
Stable Diffusion 3	sd3	✅❓	❌	✅❓	✅	✅	✅❓

❓: Untested

Host Arguments and Switches:

Argument (--argument="value")	Type	Usage
height	int	Image height
width	int	Image width
warm_up_steps	int	Number of warm up steps
port	int	Port to run Flask server
type	str	Model type name
variant	str	Torch variant
scheduler	str	Scheduler
quantize_to	str	Quantize the model
checkpoint	str	Checkpoint path
gguf_model	str	GGUF model path (usually for the UNet/Transformer)
motion_module	str	Motion module path
motion_adapter	str	Motion adapter path
control_net	str	ControlNet path
vae	str	VAE path
lora	str (dict)	LoRAs, in the format of { "path": lora_scale }
ip_adapter	str (dict)	IPAdapter, in the format of { "path": ip_scale }

Switches (--switch)	Usage
compel	Enable Compel
enable_vae_tiling	Enable VAE tiling
enable_vae_slicing	Enable VAE slicing
xformers_efficient	Enable xFormers efficient operations
enable_model_cpu_offload	Enable CPU model offloading
enable_sequential_cpu_offload	Enable CPU sequential offloading
compile_unet	Compile the UNet
compile_vae	Compile the VAE
compile_text_encoder	Compile the Text Encoder

Usage:

Check out multigpu_diffusion_comfyui and multigpu_diffusion_localai for running implementations.

Currently Implemented Software:

• AsyncDiff
• xDiT