Repository navigation

Faster prefill, tokenizer fix, draft-model speculation (1.3.0) - #9

Merged

jkyberneees merged 16 commits into

mainfrom

perf/next-pass

Nov 2, 2026

Merged

Faster prefill, tokenizer fix, draft-model speculation (1.3.0)#9
jkyberneees merged 16 commits into
mainfrom
perf/next-pass

Commits on Nov 2, 2026
- profile(): warm-up and best-of-3 timing; MLXR_BENCH_NOPROFILE
  
  jkyberneees
  and
  claude
  committed
  221ea94
- Prefill: 64x64 quantized GEMM tiles and a faster flash_prefill (bit-identical)
  
  jkyberneees
  and
  claude
  committed
  02e8089
- WIP: qmvx4 small-batch 4-bit matmul on simdgroup matrices (T = 2..8)
  
  jkyberneees
  and
  claude
  committed
  e4390d3
- WIP: scalar 2-token kernel, per-token decode attention for T <= 8, long-context verify bench
  
  jkyberneees
  and
  claude
  committed
  fa1ad8e
- fix bench main.swift
  jkyberneees
  committed
  48e073d
- qmvx4: document the measured ALU bound
  
  jkyberneees
  and
  claude
  committed
  3217d4c
- Test: a verify batch matches incremental decode (T = 2...8, plain and split attention)
  
  jkyberneees
  and
  claude
  committed
  8eaf3f3
- Draft-model speculative decoding (engine, CLI --draft, server usage)
  
  jkyberneees
  and
  claude
  committed
  269f756
- Speculative decoding tests: exact acceptance on synthetic logits, self-draft greedy identity
  
  jkyberneees
  and
  claude
  committed
  a2983ee
- App: Draft model picker (Off / Auto / compatible installed models); README
  
  jkyberneees
  and
  claude
  committed
  dfc1edb
- beginSteps: document its buffers against the in-flight limit
  
  jkyberneees
  and
  claude
  committed
  11ca590
- 1.3.0: README performance table, CHANGELOG, version
  
  jkyberneees
  and
  claude
  committed
  52da4cf
- Tokenizer: compare vocabulary keys by exact text
  
  jkyberneees
  and
  claude
  committed
  9f9d023
- Draft model: review fixes
  
  jkyberneees
  and
  claude
  committed
  437b29a
- Chat: keep message action buttons visible while the pointer is on them
  
  jkyberneees
  and
  claude
  committed
  f8aeef9
- Docs: 1.3.0 numbers from one measured set; corrected mlx-lm comparison
  
  jkyberneees
  and
  claude
  committed
  6b19940

Commits

Commits on Nov 2, 2026