fused-dense-lib / README.md
gueraf's picture
full build matrix: torch2.9-2.12 x cu126-132, x86_64+aarch64 (fused-dense-matrix-1)
e1c3364 verified
|
Raw
History Blame Contribute Delete
781 Bytes
---
tags:
- kernel
---
# fused_dense_lib
Pre-built `fused_dense_lib` CUDA extension from `csrc/fused_dense_lib` of the
[odysseyml/flash-attention](https://github.com/odysseyml/flash-attention) fork.
Backs `flash_attn.modules.mlp.FusedMLP`: `linear_bias_wgrad`, `linear_act_forward`, `bias_act_linear_dgrad_bgrad`.
No upstream wheels or community kernel exist for this extension, so the full
torch/CUDA matrix is built by the fork's `build-fused-dense-matrix.yml` workflow
(release `fused-dense-matrix-1`) and repackaged here in kernels layout.
Build variants: torch 2.9–2.12 × cu126–cu132, x86_64 + aarch64, cp312.
sm80/90 everywhere; +sm100/sm120 on cu128+.
```python
from kernels import get_kernel
fused_dense_lib = get_kernel("odyssey-systems/fused-dense-lib")
```