Prebuilt GPU kernel wheels for imbik-v2.
This repo exists only to host binaries. Upstream publishes no wheels for our
torch/CUDA line, and uv lock needs a credential-free URL to hash, which a
private repo's release assets cannot provide (they are API-only). Hence a public
host for artifacts that are, in any case, just compiled BSD-3 sources.
Built for: torch 2.11.0+cu130, CUDA 13.0, CPython 3.12, cxx11abi=TRUE. Runs on Hopper (H100/H200, sm_90) and datacenter Blackwell (B200, sm_100).
| wheel | source | GPU code | verified against |
|---|---|---|---|
flash_attn-2.8.3.post1 |
Dao-AILab/flash-attention @ v2.8.3.post1 (BSD-3) |
sm_90, sm_100 | scaled_dot_product_attention, bf16, causal, (2, 512, 8, 64): max abs diff 1.95e-03 on H200, 7.81e-03 on B200 |
causal_conv1d-1.6.2.post1 |
Dao-AILab/causal-conv1d @ v1.6.2.post1 (BSD-3) |
sm_75 to sm_121 (every arch CUDA 13.0 knows) | exact match vs grouped conv1d on a padded input, on H200 and B200 |
The flash-attn wheel is arch-specific to keep its size and build time down. To
add an arch, rebuild with a wider FLASH_ATTN_CUDA_ARCHS (see build.md).
The causal-conv1d wheel is the same binary as in kernels-cu13-torch2.11: its
setup.py picks archs from the CUDA toolkit version and ignores
TORCH_CUDA_ARCH_LIST, so the first build already covered Blackwell.
Check what a wheel contains with cuobjdump --list-elf <extension>.so.
Same sources and toolchain. Its flash_attn wheel carries sm_90 only and fails
on B200 with no kernel image is available for execution on the device.
Rebuild commands are in build.md.