| tags: | |
| - kernel | |
| # dropout_layer_norm | |
| Pre-built `dropout_layer_norm` CUDA extension from `csrc/layer_norm` of the | |
| [odysseyml/flash-attention](https://github.com/odysseyml/flash-attention) fork | |
| (tag `odyssey-v2.8.3-fused-1`), repackaged in HF kernels build-variant layout. | |
| Provides fused (dropout + residual add + LayerNorm/RMSNorm): | |
| `dropout_add_ln_fwd`, `dropout_add_ln_bwd`, | |
| `dropout_add_ln_parallel_residual_fwd`, `dropout_add_ln_parallel_residual_bwd`. | |
| Built for torch 2.9 / cu12.8 / cp312, sm_80/90/100/120 (x86_64 + aarch64). | |
| ```python | |
| from kernels import get_kernel | |
| dropout_layer_norm = get_kernel("odyssey-systems/dropout-layer-norm") | |
| ``` | |