MLX’s MoE layer is said to be getting 1.5x faster with better tile scheduling
A post credits better tile scheduling in grouped mm for the claimed speedup and links to a GitHub pull request.
TLDR
A user says MLX’s MoE layer is getting 1.5x faster through better tile scheduling in grouped mm. The post links to a GitHub pull request.
Combined views
13.2K
2 Sources, first seen 7h ago
133 likes