ollama/x
Jesse Gross 05e0f21bec mlx: fuse sigmoid router head in glm4_moe_lite
DeepSeek-V2-style aux-loss-free routing computes sigmoid(gates) once but
needs it twice: the raw sigmoid output is gathered after top-k, while the
post-bias negation is the argpartition key. Fuse into a single multi-output
Compiled kernel returning both, saving two launches on the routing path
per token. Exposed as a general SigmoidRouter since the same pattern is
shared across DeepSeek-V2 descendants.

Improves glm4.7 generation performance by approximately 1%.
2026-04-20 15:02:14 -07:00
..
agent x/cmd: enable web search and web fetch with flag (#13690) 2026-01-12 13:59:40 -08:00
cmd Reapply "don't require pulling stubs for cloud models" again (#14608) 2026-03-06 14:27:47 -08:00
create Keep Gemma4 router projection in source precision (#15613) 2026-04-15 15:04:23 -07:00
imagegen mlx: fix imagegen lookup (#15588) 2026-04-16 10:39:00 -07:00
mlxrunner mlx: fuse sigmoid router head in glm4_moe_lite 2026-04-20 15:02:14 -07:00
models mlx: fuse sigmoid router head in glm4_moe_lite 2026-04-20 15:02:14 -07:00
safetensors create: Clean up experimental paths, fix create from existing safetensor model (#14679) 2026-04-07 08:12:57 -07:00
server mlx: fix vision capability + min version (#15106) 2026-03-27 17:09:28 -07:00
tokenizer mlx: quantized embeddings, fast SwiGLU, and runtime fixes (#14884) 2026-03-17 11:21:38 -07:00
tools add ability to disable cloud (#14221) 2026-02-12 15:47:00 -08:00