You are viewing a plain text version of this content. The canonical link for it is here.
Posted to discuss-archive@tvm.apache.org by "spectrometerHBH (via GitHub)" <gi...@apache.org> on 2026/08/28 21:25:32 UTC
[PR] [Fix][Support] Use sbsa-linux CUDA include dir on ARM64 Linux [tvm]
spectrometerHBH opened a new pull request, #20222:
URL: https://github.com/apache/tvm/pull/20222
### Problem
`_compile_cuda_nvrtc` builds the architecture-specific CUDA include directory as `targets/{platform.machine()}-{platform.system().lower()}/include`, i.e. `targets/aarch64-linux` on ARM64.
CUDA toolkits for ARM64 **servers** (SBSA — GH200/GB200 and similar) install their headers under `targets/sbsa-linux`; `targets/aarch64-linux` only exists in the embedded/L4T toolkits. On an SBSA machine the computed path therefore does not exist, so both the architecture-specific include directory and the CCCL include directory derived from it are silently dropped from the NVRTC options. When the toolkit has no top-level `include/` either, compilation fails with "Cannot find CUDA headers".
### Fix
Add `_find_cuda_target_include()`, which probes `sbsa-linux` first on ARM64 Linux (`aarch64`/`arm64`) and falls back to `aarch64-linux` for embedded/L4T layouts. Other platforms keep the previous `{machine}-{system}` behavior. The helper returns `None` when no `targets/` layout is present, which is also a small cleanup of the CCCL lookup (it no longer re-runs `os.path.isdir` on the same directory).
This mirrors the CUDA target selection used by other projects on SBSA hosts, e.g. [Triton-distributed's `build_scm.sh`](https://github.com/ByteDance-Seed/Triton-distributed/blob/8260bc34398c2b8f36dc840fd22f741ca9294584/build_scm.sh#L21).
### Testing
New `tests/python/support/test_nvcc.py` covers the SBSA, L4T, and x86_64 layouts plus the no-`targets/` case against a fake toolkit tree; it runs on any host since the platform is monkeypatched.
```
python -m pytest tests/python/support/test_nvcc.py -q # 5 passed
pre-commit run --files python/tvm/support/nvcc.py tests/python/support/test_nvcc.py
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org
Re: [PR] [Fix][Support] Use sbsa-linux CUDA include dir on ARM64 Linux [tvm]
Posted by "jinhongyii (via GitHub)" <gi...@apache.org>.
jinhongyii merged PR #20222:
URL: https://github.com/apache/tvm/pull/20222
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org