Repository navigation
Would A100 / H100 validation be useful? #262
Replies: 1 comment
|
The README's A100 row is about the native build. At If you'd like to contribute an A100 result now, one bounded check is the existing FP16 causal attention test. After following the editable-install and dependency steps, it can be run independently of cmake -S . -B build-sm80 -DGPU_ARCH=80
cmake --build build-sm80 --target flash_rt_fa2 -j2
FLASHRT_BUILD_DIR="$PWD/build-sm80" python -m pytest -q -rs tests/test_fa2_fp16_causal.pyInclude the CUDA and PyTorch versions and whether the tests passed, failed, or skipped. The test can skip if the extension or entry point is missing, so that distinction is useful. SM90 is missing from that FA2 build gate too, so I'd keep the H100 result separate. This is based on a source check at your commit. Maintainers can confirm which model path they'd like to validate first. |
Uh oh!
There was an error while loading. Please reload this page.
Hi, thanks for the project.
I have access to clusters with A100 and H100 nodes. The README lists A100 as a build target, but on a real A100-SXM4-40GB (main @ cdf75a4)
detect_arch()raises "unsupported GPU SM 8.0". From the code, H100 (cc 9.0) would hit the same check.Is supporting these cards something the project wants? If so, I'd be glad to help test on them :)
All reactions