AI_Trellis2-WHLs-RTX-PRO-6000

⚡ Custom Wheels for RTX 6000 (Blackwell) - ComfyUI Trellis2

This repository documents pre-compiled Windows Wheels (.whl) optimized specifically for the NVIDIA RTX PRO 6000 Blackwell Workstation Edition. These wheels satisfy the hard-to-build dependencies required for Trellis2 implementation in ComfyUI.

These binaries are built to accelerate 3D generation and rendering pipelines, specifically targeting Compute Capability 12.0.

📥 Download

All .whl files are hosted on Hugging Face: 👉 Click here to download files from Hugging Face

⚠️ Hardware & Software Requirements

Do not install these wheels if you do not match the environment below. These are built for next-generation architecture and will likely fail on Ada Lovelace (4090) or Ampere (3090) cards due to the specific sm120 compilation.

📦 Included Wheels

The following libraries have been compiled:

1. Flash Attention (flash_attn-2.8.3)
2. Nvdiffrast (nvdiffrast-0.4.0)
3. FlexGEMM (flex_gemm-0.0.1)
4. CuMesh (cumesh-0.0.1)
5. O-Voxel (o_voxel-0.0.1)
6. Nvdiffrec (nvdiffrec-0.1.0)

🚀 ComfyUI Compatibility

These wheels are essential for running the Z-Trellis2 workflow. They have been tested with the following custom nodes:

• Recommended (Faster): visualbruno/ComfyUI-Trellis2
• Compatible: PozzettiAndrea/ComfyUI-TRELLIS2

Note: Initial testing indicates better performance/speed when using the visualbruno implementation with these specific binaries.

🛠️ Installation

This guide assumes you are using the standard ComfyUI GitHub version (cloned via git) on Windows 11.

Recommendation: It is highly recommended to install these wheels inside a Python Virtual Environment (venv) to ensure these specific CUDA-compiled wheels do not conflict with your system packages or other projects.

Step 1: Install Wheels

1. Download the .whl files from the Hugging Face Repository to a local folder (e.g., C:\wheels).
2. Open your terminal and ensure your virtual environment is active.
3. Run the following commands:

pip install "C:\wheels\flash_attn-2.8.3+cu130torch2.10.0cxx11abiTRUE-cp312-cp312-win_amd64.whl"
pip install "C:\wheels\nvdiffrast-0.4.0-cp312-cp312-win_amd64.whl"
pip install "C:\wheels\flex_gemm-0.0.1-cp312-cp312-win_amd64.whl"
pip install "C:\wheels\cumesh-0.0.1-cp312-cp312-win_amd64.whl"
pip install "C:\wheels\o_voxel-0.0.1-cp312-cp312-win_amd64.whl"
pip install "C:\wheels\nvdiffrec-0.1.0-cp312-cp312-win_amd64.whl"

Step 2: Apply Name Conversion Shim (Crucial)

The nvdiffrec wheel requires a specific module name (nvdiffrec_render) that differs from the package name. To ensure the Trellis nodes can find the module, run this Python command immediately after installation:

python -c "import site; import os; path = site.getsitepackages()[1]; f = open(os.path.join(path, 'nvdiffrec_render.py'), 'w'); f.write('from render import renderutils as _ru\nimport sys\nsys.modules[\"nvdiffrec_render\"] = _ru\nfrom render.renderutils import *'); f.close(); print('Shim created at: ' + path)"

This script creates a redirection file in your site-packages so import nvdiffrec_render works correctly.

🔗 Credits & Source Code

These wheels represent compiled versions of open-source libraries. Full credit goes to the original authors and researchers:

• Flash Attention: Tri Dao
• Other AI WHLs wildminder
• Nvdiffrast / Nvdiffrec: NVLabs
• FlexGEMM: JeffreyXiang
• CuMesh: JeffreyXiang
• TRELLIS 2 / O-Voxel: Microsoft

---

Disclaimer: These files are provided "as is" for experimental builds on Blackwell architecture. Please check the original repositories for licensing information regarding commercial use.