Fix installation issues in setup.sh#20
Conversation
- Add CUDA_HOME auto-detection with fallback to common paths - Fix linker errors by setting LIBRARY_PATH and LDFLAGS for libcuda.so - Convert git submodule URLs from SSH to HTTPS to avoid auth issues - Fix o-voxel dependencies to prevent reinstalling cumesh/flex_gemm - Add informative logging messages during installation - Update README with troubleshooting notes Fixes: - 'ld: cannot find -lcuda' linker errors during nvdiffrec/cumesh build - 'Permission denied (publickey)' errors when cloning CuMesh submodules - Dependency conflicts when installing o-voxel Co-Authored-By: Warp <agent@warp.dev>
|
Need to add |
@microsoft-github-policy-service agree |
This ensures flash-attn builds correctly with the existing environment.
|
FYI, gentoo seems to put cuda to |
The dependency entries in o-voxel/pyproject.toml live inside a quoted TOML
array, like:
dependencies = [
"cumesh @ git+https://github.com/JeffreyXiang/CuMesh.git",
...
]
The previous sed replaced `cumesh @ git+...` with `"cumesh"` (literal quotes
in the replacement), producing `""cumesh""` — invalid TOML. pip then fails
with `tomllib.TOMLDecodeError: Unclosed array`.
Strip the quotes from the replacement so the outer pair is preserved.
….7.3 flash-attn 2.7.3 has no prebuilt wheel for torch 2.7 (release ships 2.1-2.6 + 2.8 only), and source-building against torch 2.7.1+cu128 fails under modern setuptools/wheel during egg-info collection. 2.8.3 publishes an exact `cu12torch2.7cxx11abiFALSE` wheel for cp311 (and other cp3x). The bumped version is API-compatible for the dense and varlen attention paths TRELLIS.2 uses.
Summary
This PR fixes several critical installation issues in the setup.sh script that prevent successful installation of TRELLIS.2 dependencies.
Issues Fixed
1. CUDA Linker Errors
Problem: Build fails with
ld: cannot find -lcuda: No such file or directorywhen compiling nvdiffrec and cumeshSolution:
/usr/lib/x86_64-linux-gnuand CUDA stubs directory2. Git Submodule SSH Authentication Failures
Problem: CuMesh submodule clone fails with
Permission denied (publickey)errorSolution:
3. CUDA_HOME Not Set
Problem: Script assumes CUDA_HOME is set, causing issues when it's not
Solution:
/usr/local/cuda-12.4first, then/usr/local/cuda4. o-voxel Dependency Conflicts
Problem: o-voxel tries to reinstall cumesh and flex_gemm from git, causing failures
Solution:
Changes Made
setup.sh
README.md
Testing
Tested on Ubuntu with:
All packages now install successfully without manual intervention.
Related Issues
This addresses installation failures that users commonly encounter when following the setup instructions, particularly on systems with: