Conifer · Research
Research
Bench notes.
Ledger02 entries
- 01Conifer Bench · KernelsWhere Conifer beats MLX on Apple silicon, and where MLX still winsA decode-attention rewrite on an M3 Max: up to 2.9x over the prior kernel, up to 75 percent over llama.cpp, an honest loss to MLX.
- 02Conifer Bench · MemoryFitting a 24B model to a 36 GB Mac without thrashingThe KV cache, not the weights, is the limit: a 24B's full window reserves 21.5 GB on a 36 GB Mac. The fix budgets the cache and back-solves the window.