RAY
One worker
per GPU.
Ray assigns each worker to a device and coordinates distributed ComfyUI execution.
Open-source ComfyUI parallelism
RAYLIGHTGPU[S]One Workflow
Raylight splits generative AI inference across GPUs through ComfyUI's user-friendly interface.
CUDA · ROCm · XPUScroll to split the workload ↓
01 / THE PROBLEM
Just need P2P-capable GPUs.*
Raylight coordinates two or more GPUs inside ComfyUI. They do not need to be flagship cards; use the hardware that fits your budget.
BASE REQUIREMENT2+ GPUs with P2P support* Open GPU kernel modules required for non-datacenter NVIDIA GPUs.
02 / TECHNICAL CORE
Split.
Choose the parallelism that matches your bottleneck, or combine them.
One worker
USPSEQUENCE
ABCD
Split the sequence.
Ulysses and Ring parallelism divide image and video tensors among ranks, then gather the result.
FSDPWEIGHTS
Split the model.
Shard parameters when the model is larger than a single GPU’s VRAM. Add CPU offload when necessary.
UNIFIEDEXPERIMENTAL
USP × DP × CFG × FSDP
Compose parallel groups for throughput, memory capacity, or both.
CFGGUIDANCE
Split the guidance.
Run conditional and unconditional predictions on separate workers, then combine the guidance.
03 / VISUAL PROOF
See every GPU work.
An introduction to Raylight’s true multi-GPU inference workflow in ComfyUI.
RTX 2000 Ada ≈ RTX 4060 Ti 16GB. Repository figures; averages use warmed-up runs where noted.
WORKLOAD1 GPU2 GPUsRESULT
MiniMax H3864×480 · 5s + audio65s35s1.86×
Wan 2.1 14B480×480×33 · FP89.23 s/it5.18 s/it1.78×
Flux1024² · FP82.22 s/it1.26 s/it1.76×
Qwen Image1024² · FP8 + FSDPOOM5.68 s/itFits
04 / COVERAGE
From pixels to motion.
Select a model family with dedicated Raylight USP, FSDP and CFG paths.
0112
Video + audio
MiniMax H3
Native audio-video generation with synchronized dialogue, ambient sound and cinematic motion in one model path.
BYMiniMax
USPFSDP
01 / 12Video + audio
MiniMax
Video showcaseMiniMax H3Generated model output
MiniMax H3MiniMax
Native audio-video generation with synchronized dialogue, ambient sound and cinematic motion in one model path.
USPFSDP
NVIDIAVOLTA · TURING · AMPERE · ADA · BLACKWELLTested across modern CUDA architectures
AMDMI300X · MI210ROCm paths confirmed by users and maintainer testing
INTELARC PRO B60 Via Intel LLM Scaler05 / WHY COMFYUI
RAYLIGHTKeep the graph.
Split the work.
Parallel Diffusers already exists. Why use ComfyUI?
VISUAL CONTROL
Raylight brings parallel execution into an editable node graph. Loaders, conditioning, LoRAs, ControlNet and model-specific controls stay visible and composable.
MEMORY PATH
Raylight can shard model weights while ComfyUI manages and unloads workflow components, reducing overall VRAM overhead versus a comparable parallel Diffusers pipeline.
READY TO SCALE?
Your spare GPU
has a job now.
Install Raylight as a ComfyUI custom node and start with the example workflows.
git clone https://github.com/komikndr/raylight