AITER is AMD’s centralized repository that support various of high performance AI operators for AI workloads acceleration, where a good unified place for all the customer operator-level requests, which can match different customers' needs. Developers can focus on operators, and let the customers integrate this op collection into their own private/public/whatever framework.
Some summary of the features:
- C++ level API
- Python level API
- The underneath kernel could from triton/ck/asm
- Not only inference kernels, but also training kernels and gemm+comm kerenls (so can do any kerne+framework dirty WAs for any arch limit)
git clone --recursive https://github.com/ROCm/aiter.git
or
git submodule sync ; git submodule update --init --recursive
under aiter root dir run: python3 setup.py develop
there are number of op test, you can run them like this: python3 op_tests/test_layernorm2d.py
Ops | Description |
---|---|
GEMM | D=AxB+C |
FusedMoE | bf16 balabala |
WIP | coming soon... |