SIGN IN SIGN UP

FlashInfer: Kernel Library for LLM Serving

0 0 188 Python

ci(jit-cache): limit sm110 builds to aarch64 (#3275)

## Summary

- Removes `11.0a` / `sm110` from CUDA 13.0 x86_64 flashinfer-jit-cache
build arch lists.
- Keeps `11.0a` for CUDA 13.0 aarch64 builds, where Jetson AGX Thor /
T5000 targets live.
- Mirrors the same architecture split in the PR AOT build/import script.
- Fixes sccache anonymous S3 setup by clearing empty AWS credential
environment variables before setting `SCCACHE_S3_NO_CREDENTIALS=true`.

## Testing

- Parsed `release.yml`, `nightly-release.yml`, and `pr-test.yml` with
`yaml.safe_load`.
- Ran `bash -n scripts/task_test_jit_cache_package_build_import.sh`.
- Ran `bash -n scripts/jit_cache_build_common.sh
scripts/task_test_jit_cache_package_build_import.sh
scripts/build_flashinfer_jit_cache_whl.sh`.
- Mocked `setup_sccache` with empty AWS credential variables and
verified read-only anonymous mode unsets them before starting sccache.
- Mocked `setup_sccache` with non-empty AWS credentials and verified
read-write mode leaves `SCCACHE_S3_NO_CREDENTIALS` unset.
- Verified CUDA 13 architecture detection normalizes uppercase
`AARCH64`, `ARM64`, and `X86_64` inputs.
- Ran `git diff --check`.
- Verified generated CUDA 13 arch lists:
  - `x86_64`: `7.5 8.0 8.9 9.0a 10.0a 10.3a 12.0f`
  - `aarch64` / `arm64`: `7.5 8.0 8.9 9.0a 10.0a 10.3a 11.0a 12.0f`


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Chores**
* Adjusted CUDA architecture selection to choose ARM-specific CUDA
targets on aarch64 hosts, and a different list on non-ARM hosts, across
release and nightly build workflows and test scripts.
* Improved build-cache credential handling: read-write mode now requires
explicit AWS credentials; otherwise the workflow falls back to read-only
mode.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: dierksen <dierksen@dierksen-spark.localdomain>
Co-authored-by: Alex Yang <aleyang@nvidia.com>
J
Jonathan Dierksen committed
0a128d112af75de403769dd4e53edcfcc165afb9
Parent: f6717ff
Committed by GitHub <noreply@github.com> on 5/9/2026, 3:27:11 PM