Add MLX backend support for Gemma 4 31B#19524

Merged

mergennachin merged 13 commits into

mainfrom

May 19, 2026

Commits on May 19, 2026

Add MLX backend support for Gemma 4 31B
mergennachin
committed
Use shared logits_to_token sampler, support Int4 embeddings, harden MLX packer

mergennachin
and
claude
committed
Last-logits-only, MLX custom ops, remove temperature=None path, unify temp handling
mergennachin
committed
Fix MLX RoPE for proportional partial rotary (Gemma 4 full-attention layers)

mergennachin
and
claude
committed
Remove _build_masks from MLX export path for clean single-delegate graph

mergennachin
and
claude
committed
Make mlx_source_transformations self-contained

mergennachin
and
claude
committed
main.cpp cleanup + add MLX-ops graph test

mergennachin
and
claude
committed
Switch runner to Module::LoadMode::Mmap

mergennachin
and
claude
committed
Use MmapUseMlockIgnoreErrors to pin model pages in RAM

mergennachin
and
claude
committed
Fix EOS
mergennachin
committed
Restore error handling for CUDA weight_sharing_across_methods set_option

mergennachin
and
claude
committed
Fix CI config
mergennachin
committed
Unify pack_linear_for_mlx and pack_embedding_for_mlx into pack_for_mlx

mergennachin
and
claude
committed