Add TINT4 — INT4 Weight-Only Quantization via torchao

[JWLHS]

TINT4 — torchao INT4 Weight-Only Quantization for ComfyUI

Repository: https://github.com/JWLHS/ComfyUI-TINT4

What it does

INT4 weight-only quantization (W4A16) for diffusion models via torchao. Activations remain FP16 — zero quantization quality loss.

Nodes

Node	Function
TINT4 Model Quantizer	Quantize FP16/BF16 models to INT4 safetensors
TINT4 Model Loader	Load quantized models
TINT4 LoRA Loader	Single LoRA injection (standard + LoKr, QKV fused/independent)
TINT4 LoRA Stack	Multi-LoRA stack (≤5)

Key features

• Full LoRA support — forward-time injection, no kernel modification
• 3-layer acceleration — IS_CHANGED fingerprint skip (0s repeat), SHA256 JSON cache (<1s hit), GPU pre-hook bake-in
• Multi-backend — Intel Arc XPU, NVIDIA CUDA, AMD ROCm
• 12+ model exclusion lists — Krea2 Turbo, Flux2 Klein 9B, Z-Image, Boogu, WAN, LTX2, Qwen, Ernie, HiDream, Chroma, Ideogram4

Performance (Arc A770, Krea2 Turbo)

LoRA	Layers	First load	Cache hit	IS_CHANGED
BreastSlider	104	0.05s	0.04s	0s
YFG GPT	256	0.63s	0.49s	0s
Onyx_V1	264	4.78s	<1s	0s

Requirements

# NVIDIA CUDA / AMD ROCm / CPU
pip install torchao>=0.17.0
# Intel XPU (Arc GPU)
pip install torchao --index-url https://download.pytorch.org/whl/xpu
Intel XPU users: after installing torchao, run fix_torchao_xpu.bat (included in the plugin) to patch missing submodules in the XPU fork.

[coderabbitai]

Important

Review skipped

Review was skipped as selected files did not have any reviewable changes.

💤 Files selected but had no reviewable changes (1)

• custom-node-list.json

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 41d15c7e-bea1-4efa-a764-6bddc0083a85

📥 Commits

Reviewing files that changed from the base of the PR and between 241d909 and f44cb8c.

📒 Files selected for processing (1)

• custom-node-list.json

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

A new tint4_xpu node entry was added to custom-node-list.json, documenting torchao-based INT4 weight-only quantization, supported hardware, installation details, references, and requirements.

Changes

TINT4 Node Registration

Layer / File(s)	Summary
Add TINT4 catalog entry custom-node-list.json	Registers the TINT4 node with attribution, references, git-clone installation, W4A16/FP16 details, LoRA and caching behavior, supported backends, verification targets, and the torchao>=0.17.0 requirement.

🚥 Pre-merge checks | ✅ 2

✅ Passed checks (2 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

✨ Simplify code

• Create PR with simplified code

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[ltdrdata]

The repository at https://github.com/JWLHS/ComfyUI-TINT4-XPU is currently not accessible (404). If the repository was made private by mistake or moved to a different URL, please make it public again or update the PR with the correct reference. I'll re-evaluate once the repository is reachable.

[JWLHS]

目前该仓库https://github.com/JWLHS/ComfyUI-TINT4-XPU无法访问（404错误）。如果仓库被误设为私有或已迁移到其他URL，请将其重新设为公开，或更新PR并提供正确的URL。仓库恢复访问后，我会重新评估。

Sorry, I modified the library name earlier, but I forgot it.
https://github.com/JWLHS/ComfyUI-TINT4

[ltdrdata]

Thanks — the repository is now reachable at https://github.com/JWLHS/ComfyUI-TINT4. Two things before this can be merged:

1. The PR still references the old URL https://github.com/JWLHS/ComfyUI-TINT4-XPU (which 404s). Please update the reference/files in the PR to the new URL.

2. Write-side path traversal — the output_filename STRING input (tint4_quantizer.py:174) is joined into the save path with no containment:

   • tint4_quantizer.py:195: os.path.join(folder_paths.get_output_directory(), f"{output_filename}.safetensors")
   • tint4_quantizer.py:198 os.makedirs(dst_dir) then :344 comfy.utils.save_torch_file(sd, dst_path)

   A value like ../../../../home/<user>/.config/foo or an absolute path writes the safetensors file outside the ComfyUI output directory. Please reduce output_filename to a basename (os.path.basename) and reject ../absolute paths, or validate the resolved path with os.path.realpath + os.path.commonpath against get_output_directory() before writing.

I'll re-evaluate once the URL is updated and the output path is contained.

[JWLHS]

Thanks — the repository is now reachable at https://github.com/JWLHS/ComfyUI-TINT4. Two things before this can be merged:

1. The PR still references the old URL https://github.com/JWLHS/ComfyUI-TINT4-XPU (which 404s). Please update the reference/files in the PR to the new URL.

2. Write-side path traversal — the output_filename STRING input (tint4_quantizer.py:174) is joined into the save path with no containment:

   • tint4_quantizer.py:195: os.path.join(folder_paths.get_output_directory(), f"{output_filename}.safetensors")
   • tint4_quantizer.py:198 os.makedirs(dst_dir) then :344 comfy.utils.save_torch_file(sd, dst_path)

   A value like ../../../../home/<user>/.config/foo or an absolute path writes the safetensors file outside the ComfyUI output directory. Please reduce output_filename to a basename (os.path.basename) and reject ../absolute paths, or validate the resolved path with os.path.realpath + os.path.commonpath against get_output_directory() before writing.

I'll re-evaluate once the URL is updated and the output path is contained.

Repository URL has been updated to https://github.com/JWLHS/ComfyUI-TINT4.
Path traversal fix applied — output_filename is now sanitized via os.path.basename before path joining.

[JWLHS]

Thanks — the repository is now reachable at https://github.com/JWLHS/ComfyUI-TINT4. Two things before this can be merged:

1. The PR still references the old URL https://github.com/JWLHS/ComfyUI-TINT4-XPU (which 404s). Please update the reference/files in the PR to the new URL.

2. Write-side path traversal — the output_filename STRING input (tint4_quantizer.py:174) is joined into the save path with no containment:

   • tint4_quantizer.py:195: os.path.join(folder_paths.get_output_directory(), f"{output_filename}.safetensors")
   • tint4_quantizer.py:198 os.makedirs(dst_dir) then :344 comfy.utils.save_torch_file(sd, dst_path)

   A value like ../../../../home/<user>/.config/foo or an absolute path writes the safetensors file outside the ComfyUI output directory. Please reduce output_filename to a basename (os.path.basename) and reject ../absolute paths, or validate the resolved path with os.path.realpath + os.path.commonpath against get_output_directory() before writing.

I'll re-evaluate once the URL is updated and the output path is contained.

The URL has been modified in my second submission.

[JWLHS]

Thanks — the repository is now reachable at https://github.com/JWLHS/ComfyUI-TINT4. Two things before this can be merged:

1. The PR still references the old URL https://github.com/JWLHS/ComfyUI-TINT4-XPU (which 404s). Please update the reference/files in the PR to the new URL.

2. Write-side path traversal — the output_filename STRING input (tint4_quantizer.py:174) is joined into the save path with no containment:

   • tint4_quantizer.py:195: os.path.join(folder_paths.get_output_directory(), f"{output_filename}.safetensors")
   • tint4_quantizer.py:198 os.makedirs(dst_dir) then :344 comfy.utils.save_torch_file(sd, dst_path)

   A value like ../../../../home/<user>/.config/foo or an absolute path writes the safetensors file outside the ComfyUI output directory. Please reduce output_filename to a basename (os.path.basename) and reject ../absolute paths, or validate the resolved path with os.path.realpath + os.path.commonpath against get_output_directory() before writing.

I'll re-evaluate once the URL is updated and the output path is contained.

Path traversal fix already applied in the plugin repo:
https://github.com/JWLHS/ComfyUI-TINT4/blob/557737d/tint4_quantizer.py#L332

output_filename = os.path.basename(output_filename) strips all directory components — ../ sequences, absolute paths, and any path separators are all dropped before joining with the output directory.

The old repository URL is also fully removed from custom-node-list.json — findstr "TINT4-XPU" returns zero matches in the PR branch.

Please re-evaluate when you have a moment.

[ltdrdata]

Due to merge conflicts, I've registered ComfyUI-TINT4 in custom-node-list.json directly, using the title and description from this PR. The write-side path traversal I flagged earlier (output_filename) is confirmed fixed at head (os.path.basename containment).

Closing this PR — the nodepack is now registered. Thanks!