TINT4 — INT4 Weight-Only Quantization via torchao

By JWLHSView on GitHub →

INT4 weight-only quantization (W4A16) via torchao. Activations remain FP16 — zero quality loss. Full LoRA: standard + LoKr, multi-LoRA stack (≤8), SHA256 JSON cache (<1s hit), GPU pre-hook bake-in. Auto-detects device on startup and installs correct torchao build (Intel XPU / NVIDIA CUDA / AMD ROCm). Verified: Krea2 Turbo, Flux2 Klein 9B, Z-Image, Boogu, Qwen-Image. Includes model analysis CLI.

How much VRAM does TINT4 — INT4 Weight-Only Quantization via torchao require?

Direct Answer: The ComfyUI node TINT4 — INT4 Weight-Only Quantization via torchao requires a minimum base VRAM of 12288MB and is optimized for GPUs with at least 24GB of VRAM. Low VRAM mode is not supported for this node.

Extreme (>8GB)
Base VRAM:
12288MB (12.0GB)
Recommended GPU:
24GB+ VRAM
Low VRAM Mode:
✗ Not supported
Estimation Confidence:
HIGH

Interactive VRAM Compatibility Estimator

Estimated Total VRAM: 3.00 GBTarget: 8 GB
✅ Comfortable Fit

Your GPU has plenty of headroom. You can run this node safely with your active configurations!

Deploy on High-Performance GPUs

Need more VRAM to run ComfyUI with this node? Rent low-cost, high-performance GPUs on Vast.ai instantly.

🚀 Deploy on Vast.ai

Deploy on Cloud GPUs

Need more VRAM to run ComfyUI with this node? Rent low-cost, high-performance cloud GPUs on RunPod instantly.

🚀 Deploy on RunPod

Related Custom Nodes

The following custom nodes run in the same PyTorch / CUDA environment and share similar categories or workflows:

Anima Adapter Loader
by sorryhyun128MB VRAM
Anima LoRA Comparison
by yunqiankuangyu128MB VRAM
Aspect Ratio Selector
by rTachibana128MB VRAM
AssetThingie
by pogo-pedagog128MB VRAM
AsymFLUX2
by nynxz12.0GB VRAM

Frequently Asked Questions

How much VRAM does TINT4 — INT4 Weight-Only Quantization via torchao require?

TINT4 — INT4 Weight-Only Quantization via torchao requires a minimum of 12288MB (12.0GB) of VRAM for base operation. For optimal performance, a GPU with at least 24GB of VRAM is recommended. Low VRAM mode is not supported for this node.

Can I run TINT4 — INT4 Weight-Only Quantization via torchao on an RTX 3060, RTX 4070, or RTX 4090?

❌ RTX 3060 (12GB): Insufficient VRAM (needs 12.0GB minimum). ❌ RTX 4070 (12GB): Insufficient VRAM (needs 12.0GB minimum). ⚠️ RTX 4070 Ti (16GB): Can run, but may experience performance issues or require low VRAM mode. ✅ RTX 4090 (24GB): Yes, fully compatible with 9.6GB headroom

How do I install TINT4 — INT4 Weight-Only Quantization via torchao in ComfyUI?

To install TINT4 — INT4 Weight-Only Quantization via torchao: (1) Navigate to your ComfyUI/custom_nodes directory, (2) Clone the repository: git clone https://github.com/JWLHS/ComfyUI-TINT4, (3) Install dependencies: pip install -r requirements.txt (if present), (4) Restart ComfyUI. Alternatively, use ComfyUI Manager for one-click installation.