Cyberdelia Quantizer

ComfyUI Cyberdelia Quantizer

<p align="center"> <img src="assets/icon.png" alt="Cyberdelia Quantizer icon" width="200"> </p>

An independent ComfyUI custom node for converting original FP16/BF16 diffusion checkpoints to native mixed-precision safetensors files.

The main compatibility fix is deliberately simple: every quantized weight gets an explicit marker beside the exact physical layer name:

model.diffusion_model.blocks.0.attn.to_q.weight
model.diffusion_model.blocks.0.attn.to_q.weight_scale
model.diffusion_model.blocks.0.attn.to_q.comfy_quant

The prefix is never stripped. This prevents the central-metadata prefix mismatch that can make Forge Neo treat an INT8 checkpoint as FP16 and report unexpected SingleStreamDiT keys.

Formats

| Option | Storage | Runtime notes | | --- | --- | --- | | FP8_E4M3 | FP8 E4M3 | Portable FP8 storage; full-precision matmul for compatibility | | FP8_E5M2 | FP8 E5M2 | Wider range, lower mantissa precision; full-precision matmul | | NVFP4 | packed 4-bit | Native acceleration on NVIDIA Blackwell (SM 10.0+) | | MXFP8 | block-scaled FP8 | Native acceleration on NVIDIA Blackwell (SM 10.0+) | | INT8 | per-channel INT8 | Native ComfyUI format; accelerated on NVIDIA Turing or newer | | INT8_CONVROT | rotated per-channel INT8 | Recommended broad option; NVIDIA Turing or newer | | INT4_CONVROT | packed rotated INT4 | Smaller; NVIDIA Turing or newer; profile-sensitive | | W4A8_INT8 | grouped 4-bit weights / INT8 runtime | Requires a recent kitchen backend; Ampere-class path or newer |

NVFP4 and MXFP8 files still load on unsupported hardware, but ComfyUI may fall back to dequantized computation instead of providing a speed-up. FP8 in this node uses the official compatibility-oriented marker full_precision_matrix_mult=true.

Requirements

• A current ComfyUI installation (version 0.34 or newer is recommended).
• The comfy-kitchen version pinned by that ComfyUI release.
• An original dense FP16/BF16 .safetensors diffusion model.
• On Windows, safetensors 0.8 or newer is recommended. The node then reads tensors with the pread backend instead of mmap, which avoids a known access violation crash when large checkpoints are loaded inside a long-running ComfyUI process. Older versions still work but fall back to mmap.

Do not manually upgrade only comfy-kitchen while leaving ComfyUI behind. Update ComfyUI and its requirements together so their quantization contracts match.

Do not use an FP8, INT8, GGUF, NVFP4, or other already-quantized checkpoint as the source. The node rejects recognized quantized inputs instead of silently applying a second lossy conversion.

Installation

ComfyUI Manager

After the first Comfy Registry release, search for Cyberdelia Quantizer in ComfyUI Manager, install it, and restart ComfyUI.

Git

From the ComfyUI custom_nodes directory:

git clone https://github.com/cyberdeliaAI/ComfyUI-Cyberdelia-Quantizer.git

Manual installation

Download a release archive and extract it as:

ComfyUI/custom_nodes/ComfyUI-Cyberdelia-Quantizer

Then update ComfyUI's normal requirements and restart ComfyUI. No separate node dependencies are installed because Torch, safetensors, and comfy-kitchen must match the host ComfyUI build.

Nodes

🍳 Cyberdelia Quantizer

1. Select the original model in model_name.
2. Select the matching model_profile.
3. Select the output quant_format.
4. Leave output_filename empty to use the source filename plus a format suffix.
5. Run once with dry_run=true if you want a plan without writing a file.
6. Keep strict=true for release builds.

Before a real conversion the node unloads the models ComfyUI has loaded, so the next generation reloads them. A dry run leaves them in place.

The output is written beside the source model. Existing files are protected unless overwrite=true. Saving uses a temporary file, validates its marker/weight/sidecar contract, and only then moves it to the final filename.

🔎 Cyberdelia Quant Check

This checks only the safetensors header and tiny JSON markers. It identifies:

• exact per-layer .comfy_quant packaging;
• legacy central _quantization_metadata;
• the model.diffusion_model.* prefix mismatch;
• orphan markers and missing side tensors.

It does not load the model weights into RAM.

Profiles

Dedicated profiles are included for MOSS-Image Base/SFT/Turbo, Krea2, Anima, Z-Image Base/Turbo, Flux 1/2, Qwen Image, Wan 2.2, LTX-2, ACE-Step, Boogu Image, Chroma, ERNIE Image, Ideogram 4, MiniMax H3, and SeedVR. A conservative generic profile is also available.

A profile controls which sensitive layers stay in their original source dtype. Shape-incompatible layers are also kept unchanged and listed in the result. The Qwen Image 2512 profile keeps its sensitive txt_mlp/txt_mod weights in FP8 E4M3 when a more aggressive format is selected.

MOSS-Image / MOSS Turbo

New in 1.1.0: choose MOSS-Image for an original dense MOSS Base, SFT, fine-tune or merged Turbo checkpoint. The same profile handles separate Q/K/V weights and fused QKV weights in the native ComfyUI MOSS layout.

The profile quantizes only the main blocks' attention and FFN projections. It preserves the following in their original source dtype:

• Image and text embedders, including both text-refiner blocks.
• Time embedding and shared block modulation.
• Output modulation, the final image projection, norms and biases.
• Other weights outside the recognized main-block projections.

Unprefixed keys, diffusion_model.*, and model.diffusion_model.* are supported; the physical prefix remains unchanged in every marker and side tensor. Unknown layouts produce no eligible main-block weights rather than treating unrecognized layers as MOSS attention weights.

For an SFT model with a separate DMD adapter, convert the dense SFT model first, then load the DMD LoRA in the generation workflow. Start with quant_format=INT8_CONVROT, strict=true and a dry run. Compare the real conversion against the dense model with matching prompt, seed and sampling settings before relying on image quality. The other formats remain available subject to their existing backend and hardware requirements.

This profile targets the diffusion model. It does not add Qwen3.5 text-encoder or VAE conversion. Header/shape tests cover the native 42-block MOSS layout; small-tensor conversion checks do not establish quality on a full checkpoint.

Design differences from the Tritant converter

• Writes exact per-layer .comfy_quant tensors; no prefix-stripped central map.
• Adds real FP8 E4M3 and FP8 E5M2 output modes (MXFP8 remains separate).
• Reads source tensors one at a time instead of loading the complete dense source checkpoint before conversion.
• Preserves excluded and non-floating tensors exactly instead of casting every unquantized tensor to BF16.
• Aborts on layer failures by default instead of returning a misleading success.
• Refuses accidental source overwrite and uses an atomic validated save.
• Chooses the largest valid ConvRot group per layer (256, 128, 64, 32, or 16).

Tests

The dependency-free contract tests can be run outside ComfyUI:

python -m unittest discover -s tests -v

They specifically guard against reintroducing the prefix-stripping bug.

License and upstream references

Apache-2.0. See NOTICE for attribution. This project is not an official Comfy Org project. The checkpoint contract follows the public ComfyUI native quantization loader and comfy-kitchen layouts.

Repository: https://github.com/cyberdeliaAI/ComfyUI-Cyberdelia-Quantizer