ComfyUI XPO3

ComfyUI nodes for the promoted XPO3 runtime stack.

The current release supports Mage-Flow-Turbo-XPO3-NVFP4, a four-step derivative of Microsoft Mage-Flow-Turbo for native Blackwell NVFP4 execution with the validated ComfyUI runtime stack bundled in this plugin.

XPO3: three precisions, one fused runtime. NVFP4 provides transformer speed, FP8 compresses the text encoder, and BF16 is retained where our matched testing found that lower precision caused visible damage.

Promoted runtime stack

The v0.3.0 ComfyUI runtime includes the validated composed stack:

• fused GELU-up on eligible NVFP4 up projections;
• calibrated FP4 image-MLP bridge on image blocks 0-8;
• official Sparge/Sage2 attention on steps 1-2 across all 12 blocks;
• the CFG1, batch-one direct-HND attention processor on that routed path;
• exact packaged fallback on steps 0 and 3;
• exact fallback when a toggle is off or the runtime envelope is unsupported.

Each feature remains independently configurable from the Mage-Flow Turbo XPO3 Generate node:

• enable_fused_gelu_up
• enable_fp4_bridge
• bridge_blocks
• enable_attention_accel
• enable_direct_hnd
• attention_steps
• attention_blocks

Fallback rules are explicit:

• disabling fused GELU-up restores the original activation module call;
• disabling the bridge, removing a bridge block, or hitting an unsupported shape keeps that block on the packaged path;
• attention routing is enabled only for four-step, CFG1 generations on the supported pipe, with Sparge available;
• steps 0 and 3 always stay on the packaged exact attention path;
• unsupported attention envelopes fall back to the original packaged route.

Model files

Place the three release files in the normal ComfyUI model directories:

ComfyUI/models/
├── diffusion_models/
│   └── Mage-Flow-Turbo-XPO3-NVFP4.safetensors
├── text_encoders/
│   └── qwen3vl_4b_fp8_scaled.safetensors
└── vae/
    └── Mage-Flow-VAE.safetensors

Component	Size
XPO3 diffusion model	4.87 GB
Scaled-FP8 Qwen3-VL text encoder	5.24 GB
Mage VAE	345 MB

Installation

Use a dedicated ComfyUI virtual environment:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/AJH-Code/ComfyUI-XPO3.git
cd ComfyUI-XPO3
python -m pip install -r requirements-runtime-cu130.txt
CUDA_HOME=/usr/local/cuda-13.1 \
  python -m pip install --no-build-isolation flash-attn==2.8.3
python -m pip install -r /path/to/ComfyUI/requirements.txt

Restart ComfyUI, then add:

1. Load Mage-Flow Turbo XPO3
2. Mage-Flow Turbo XPO3 Generate
3. Preview Image or Save Image

The starter API workflow is example_workflows/mage_flow_turbo_xpo3_api.json. The starter workflow and node default to 1280x1280. Recommended settings are four steps, CFG 1.0, and static shift 6.0.

Speed

The promoted stack was validated on a robust matrix of 3 prompts x 3 fixed seeds per GPU/resolution, with one warmup and three measured iterations per row. Table values below are candidate warm-generation medians across the nine case/seed rows at each GPU/resolution.

GPU	Resolution	Warm generation	Vs. packaged FA2 baseline
RTX 5080	1024×1024	0.5996 s	1.1862x
RTX 5080	1280×1280	0.9570 s	1.2116x
RTX 5060 Ti	1024×1024	1.3012 s	1.1771x
RTX 5060 Ti	1280×1280	2.1308 s	1.2034x

Across the matrix, end-to-end throughput ranged from 1.177x to 1.212x versus the previous packaged resident FA2 baseline. This is not a comparison to Microsoft's BF16 pipeline.

Reproduce the benchmark

benchmark.py runs the promoted node defaults through the actual ComfyUI loader/generator surface. It uses the fixed retriever, fisherman, and glass-house prompts; four steps; CFG 1; static shift 6; a fresh-cache warm-up; and repeated hot generations. From the plugin directory and the same ComfyUI Python environment:

python benchmark.py --width 1024 --height 1024 \
  --output xpo3_1024_benchmark.json

The JSON report records every prompt/seed, timing distribution, environment, load/placement time, and active-feature report. It will not overwrite an existing report. Results vary by GPU, driver, and ComfyUI environment.

Peak allocated memory increased modestly versus the packaged baseline:

• 1024: about +51.0 MiB
• 1280: about +78.6 MiB

The composed stack still stayed under 16 GB on both validated Blackwell cards.

Compatibility

The bundled runtime is an exact binary release for this environment:

Component	Required
Operating system	Linux x86-64
GPU	NVIDIA Blackwell SM120
VRAM	Under 16 GB on the validated stack; 16 GB class validated
Python	3.11
PyTorch	2.13.0+cu130
TorchVision	0.28.0+cu130
TorchAudio	2.11.0+cu130
CUDA ABI used to build bundled binaries	13.1
Diffusers	0.38.0
Transformers	5.5.0
comfy-kitchen	0.2.22
flash-attn	2.8.3

Windows, WSL, pre-Blackwell GPUs, Python 3.12/3.13, and other PyTorch/CUDA ABIs are not supported by this packaged runtime.

Current integration boundary

The weights use ComfyUI's normal diffusion-model, text-encoder, and VAE folders and dropdowns. Mage-Flow is not currently implemented by ComfyUI's generic diffusion loader, so the checkpoint still requires these XPO3 custom nodes. It does not yet connect directly to stock KSampler, LoRA, ControlNet, or native Comfy conditioning nodes.

Licenses and bundled third-party code

• XPO3 custom-node and packaging code: MIT.
• Mage-Flow-Turbo and Mage source: upstream MIT terms.
• Qwen3-VL and the scaled-FP8 text checkpoint: Apache-2.0.
• Bundled SpargeAttn runtime code and extensions under runtime/spas_sage_attn/: Apache-2.0.

The SpargeAttn license text is included at third_party_licenses/SpargeAttn-LICENSE. The upstream project is thu-ml/SpargeAttn. Additional bundled-source notes are summarized in THIRD_PARTY_NOTICES.md.