Repository navigation
Group-GEMM in XLA: Making Mixture-of-Experts Fast and Memory-Efficient — ROCm Blogs #642
giscus[bot]
Bot
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Group-GEMM in XLA: Making Mixture-of-Experts Fast and Memory-Efficient — ROCm Blogs
Learn how to make JAX/MaxText MoE training more efficient with two new OpenXLA Group-GEMM backends for ragged-dot: ROCm hipBLASLt and generic Triton.
https://rocm.blogs.amd.com/software-tools-optimization/openxla-group-gemm/README.html
All reactions