mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-04-29 12:19:50 +00:00
1.2 KiB
1.2 KiB
USAGE
- To use this feature, you should use MOE_INT8 method (i.e.
--kt-method MOE_INT8) - !!! you should see the method in the below motivation section to build and install the correct amd blis lib.
- Before your install you should set
export CPUINFER_ENABLE_BLIS=ONto enable
Motivation
To accelerate the prefill speed of AMD. Reference the https://github.com/amd/blis repo. And the usage should add the LPGEMM support. See the docs here: https://www.cs.utexas.edu/~flame/BLISRetreat2024/slides/Bhaskar_BLIS_Retreat_2024_AMD_LPGEMM_0.pdf
I reference this api guide for the code: https://docs.amd.com/r/en-US/57404-AOCL-user-guide/AOCL-BLAS?section=lpgemm-in-aocl-blas
To use lpgemm, see the doc here:
https://www.amd.com/content/dam/amd/en/documents/developer/version-4-1-documents/aocl/aocl-4-1-user-guide.pdf
So, you just need to enable aocl_gemm add-on, examples are here:https://github.com/amd/blis/blob/master/docs/CMakeBuildSystem.md