horizonrobotics Affordance Deployment
Use the repository installer first. It initializes both submodules, builds
torch-scatter, flash-attn==2.8.2, and GraspGen's pointnet2_ops, then
restores the supported OpenCV versions.
bash install.sh affordance
affordance-cli --help
img3d-cli --help
CUDA Architecture
Load the CUDA hook before rebuilding extensions and verify the selected target:
echo "$CUDA_HOME"
echo "$TORCH_CUDA_ARCH_LIST"
nvcc --version
- RTX 4090:
TORCH_CUDA_ARCH_LIST=8.9. - RTX 5090:
TORCH_CUDA_ARCH_LIST=12.0and nvcc 12.8 or newer.
If pointnet2_ops was built for the wrong architecture, rebuild the package
directly instead of running thirdparty/GraspGen/install_pointnet.sh, which
hardcodes an older target:
python -m pip install --no-build-isolation --force-reinstall --no-deps \
thirdparty/GraspGen/pointnet2_ops
Flash Attention
Keep flash-attn at 2.8.2 because the installed diffusers version rejects
newer releases. Rebuild from source with the GPU-specific target when needed:
export FLASH_ATTENTION_FORCE_BUILD=TRUE
export FLASH_ATTN_CUDA_ARCHS=89 # use 120 on RTX 5090 with nvcc >= 12.8
python -m pip install --no-cache-dir --no-build-isolation \
--no-binary flash-attn --force-reinstall --no-deps flash-attn==2.8.2
If the source checkout lacks CUTLASS or composable-kernel files, initialize its
submodules before retrying. Reduce MAX_JOBS when compilation runs out of RAM.
Quick Reference
| Error | Root Cause | Fix |
|---|---|---|
Unsupported gpu architecture 'compute_120' |
nvcc is older than 12.8 | Use the cu128 hook or target the actual GPU |
no kernel image in pointnet2 |
Extension built for another GPU | Rebuild pointnet2_ops directly |
| diffusers rejects flash-attn | Version is newer than 2.8.2 | Reinstall 2.8.2 |
| Build OOM | Too many compiler jobs | Set MAX_JOBS=4 |