[TRELLIS.2] Installation fails on Ubuntu 20.04 — nvdiffrast glibc mismatch & PyTorch 2.4 lacks Python 3.13 wheels

[bien-phillip]

---

Describe the Bug

Installing the [TRELLIS.2] Image-to-3D Node package (option 8) via installer.py fails with two separate errors during the isolated environment setup:

1. Primary attempt (CUDA 12.8 + PyTorch 2.8): The pre-built nvdiffrast wheel requires manylinux_2_34 or manylinux_2_35, but Ubuntu 20.04 provides glibc 2.31 (manylinux_2_31), making the
   wheel incompatible.
2. Fallback attempt (CUDA 12.4 + PyTorch 2.4): torch==2.4.* does not provide wheels for CPython 3.13 (cp313). Only cp38 through cp312 are available, so dependency resolution fails.

To Reproduce

1. Set up a system with Ubuntu 20.04, Python 3.13, NVIDIA driver 570.x, and CUDA 12.8.
2. Run python ./installer.py "/path/to/ComfyUI"
3. Select option 8 — [TRELLIS.2] Image-to-3D Node
4. Confirm all prompts with y
5. Installation fails at the comfy-env isolated environment build stage

Expected Behavior

The TRELLIS.2 node and its isolated Python environment (including nvdiffrast, torch, flash-attn, etc.) should install successfully.

Actual Behavior

Installation fails twice:

• First attempt errors out with:
  A URL dependency is incompatible with the current platform:
  nvdiffrast-0.4.0+cu128torch2.8-cp313-cp313-manylinux_2_34_x86_64.manylinux_2_35_x86_64.whl

hint: The wheel is compatible with Linux (manylinux_2_34_x86_64), Linux (manylinux_2_35_x86_64),
but you're on Linux (manylinux_2_31_x86_64)

• Fallback (CUDA 12.4) errors out with:
  torch>=2.4.0,<=2.4.1+cu124 has no wheels with a matching Python ABI tag (e.g., cp313)
  we can conclude that torch>=2.4.0,<=2.4.1+cu124 cannot be used.

hint: You require CPython 3.13 (cp313), but we only found wheels for torch (v2.4.1+cu124)
with the following Python ABI tags: cp38, cp39, cp310, cp311, cp312

Screenshots / Videos

N/A — console logs provided below.

Console Logs

1. ComfyUI Server Console:

[comfy-env] Selected: CUDA 12.8 + PyTorch 2.8
Installing CUDA packages from https://download.pytorch.org/whl/cu128
nvdiffrast from
https://github.com/PozzettiAndrea/cuda-wheels/releases/download/nvdiffrast-latest/nvdiffrast-0.4.0%2Bcu128torch2.8-cp313-cp313-manylinux_2_34_x86_64.manylinux_2_35_x86_64.whl
RuntimeError: Failed to install nvdiffrast:
stderr: error: Failed to determine installation plan
Caused by: A URL dependency is incompatible with the current platform
hint: The wheel is compatible with Linux (manylinux_2_34_x86_64), Linux (manylinux_2_35_x86_64), but you're on Linux (manylinux_2_31_x86_64)

--- Retry with CUDA 12.4 fallback ---

[comfy-env] Selected: CUDA 12.4 + PyTorch 2.4
RuntimeError: pixi install failed:
stderr: Error: failed to solve the pypi requirements of environment 'default' for platform 'linux-64'
Because torch>=2.4.0,<=2.4.1+cu124 has no wheels with a matching Python ABI tag (e.g., cp313),
we can conclude that torch>=2.4.0,<=2.4.1+cu124 cannot be used.
hint: You require CPython 3.13 (cp313), but we only found wheels for torch (v2.4.1+cu124) with the following Python ABI tags: cp38, cp39, cp310, cp311, cp312

System Information:

• Blender Version: N/A (running installer.py directly)
• StableGen Version: Latest (installer.py from main branch)
• Operating System: Ubuntu 20.04.2 LTS (glibc 2.31)
• ComfyUI Installation: Standard local install at /home/user/comfy/ComfyUI
• GPU: NVIDIA GeForce RTX 3080 (10 GB)
• NVIDIA Driver: 570.124.06
• CUDA Version: 12.8
• Python Version: 3.13.3 (CPython)
• comfy-env Version: 0.2.0
• ComfyUI-TRELLIS2 Pinned Commit: 6b0b2148f45b
• StableGen Installation Method: installer.py
  • Package selected: 8. [TRELLIS.2] Image-to-3D Node

StableGen Parameters (if relevant):

N/A — failure occurs during installation, before any generation parameters can be set.

Is the issue scene specific?

N/A — this is an installation-time failure, not a runtime issue.

Additional Context

The root cause appears to be a combination of two factors:

1. Ubuntu 20.04's glibc (2.31) is too old for the pre-built nvdiffrast wheel hosted at PozzettiAndrea/cuda-wheels, which targets manylinux_2_34 (glibc 2.34+). This blocks the CUDA 12.8 +
   PyTorch 2.8 path.
2. The CUDA 12.4 fallback pins PyTorch 2.4, which was released before Python 3.13 support was added (PyTorch 3.13 wheels are only available from PyTorch 2.5+). This blocks the fallback
   path entirely.

Possible solutions:

• Provide nvdiffrast wheels built against manylinux_2_31 (or manylinux_2_28) for broader OS compatibility.
• Update the CUDA 12.4 fallback to use PyTorch 2.5+ which includes cp313 wheels.
• Document Python 3.12 as the maximum supported version, or Ubuntu 22.04+ as the minimum supported OS.

[bien-phillip]

Hi. Thank you for great work. This issue accurred on windows also.

[sakalond]

I will look into it. The dependency install for the TRELLIS.2 node is very problematic and I already had to fix it a couple of times.