| library_name: kernels | |
| tags: | |
| - kernels | |
| - rocm | |
| # MLA (Multi-head Latent Attention, RDNA4/gfx1201) | |
| Multi-head latent attention decode / prefill / verify (+fp8) native HIP kernels. | |
| Built with [kernel-builder](https://github.com/huggingface/kernel-builder) for AMD RDNA4 (gfx1201). | |
| Load with the [`kernels`](https://github.com/huggingface/kernels) library: | |
| ```python | |
| from kernels import get_kernel | |
| kernel = get_kernel("pat883/mla-hip") | |
| ``` | |
| > Requires a ROCm PyTorch build (torch 2.10 / ROCm 7.x) on an RDNA4 card. Built variants: | |
| > `torch210-cxx11-rocm70` and `torch210-cxx11-rocm71`. | |