monster/lato.2_mrp_mlx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Commit Graph

  • ee86bdb4e4 Pass mode='max' explicitly to downsample main John 2026-08-02 10:42:18 +1000
  • 21c9de4bdd Depend on trellis_sparse_mlx instead of vendoring the sparse core John 2026-08-02 10:31:41 +1000
  • d48864033e V-VAE encoder running on the real checkpoint John 2026-08-02 10:22:13 +1000
  • 95fd4496da Sparse layer set: linear, norms, activations, resblock, attention, transformer blocks John 2026-08-02 10:18:38 +1000
  • db2097c19f Record fleet benchmark results and the optimisation story John 2026-08-02 10:11:00 +1000
  • 2a0f64375c Cache the prepared device gather index (5.9x total on the hot op) John 2026-08-02 10:09:57 +1000
  • 6fcf677cef Fuse the K^3 taps into one gather + matmul (2.1x on M3 Ultra) John 2026-08-02 10:07:53 +1000
  • 97dcdfb54a MLX sparse core: SubMConv3d + SparseTensor + weight converter John 2026-08-02 10:04:24 +1000