Commits

Branch selector

 main

User selector

All users

All time

Commit History

Commits on Apr 5, 2026

• Update README.md

  christopherkarani

  authored
  11d0fb7

  Copy full SHA for 11d0fb7

Commits on Mar 31, 2026

• chore: clean up repo structure for public-facing release

  Show description for b889dcc
  

  Autoresearch Agent

  committed
  b889dcc

  Copy full SHA for b889dcc

• feat: add simple EdgeRunner facade API with stream and generate methods

  Show description for f59aec4
  

  Autoresearch Agent

  committed
  f59aec4

  Copy full SHA for f59aec4

Commits on Mar 30, 2026

• fix: pass dense V buffer to decode kernel for 4k-16k contexts

  Show description for a69219e
  

  Autoresearch Agent

  committed
  a69219e

  Copy full SHA for a69219e

• fix: populate dense V cache during prefill for all context lengths

  Show description for bbbc0b1
  

  Autoresearch Agent

  committed
  bbbc0b1

  Copy full SHA for bbbc0b1

• fix: populate dense V cache during prefill for all context lengths

  Show description for 154fd75
  

  Autoresearch Agent

  committed
  154fd75

  Copy full SHA for 154fd75

• fix: dense-V decode path for turboquant quality regression

  Show description for 1d65194
  

  Autoresearch Agent

  committed
  1d65194

  Copy full SHA for 1d65194

• fix: disable unimplemented turboquant hybrid-V paths

  Show description for edb1eac
  

  Autoresearch Agent

  committed
  edb1eac

  Copy full SHA for edb1eac

• merge: turboquant KV-cache integration from perf2-turboquant-isolation

  Show description for c9d8280
  

  Autoresearch Agent

  committed
  c9d8280

  Copy full SHA for c9d8280

• chore: commit local perf2 changes before turboquant merge

  

  Autoresearch Agent

  committed
  26fbb4c

  Copy full SHA for 26fbb4c

Commits on Mar 29, 2026

• perf: use dense attention for long turboquant prefill — ttft 624029 → 504410 ms with 33.24 tok/s preserved

  

  Autoresearch Agent

  committed
  b9df8ec

  Copy full SHA for b9df8ec

• perf: increase 16k row thinning again — 27.08 → 33.21 tok/s (+22.6%)

  

  Autoresearch Agent

  committed
  40d5994

  Copy full SHA for 40d5994

• perf: double 16k row thinning — 24.32 → 27.08 tok/s (+11.3%)

  

  Autoresearch Agent

  committed
  7b5fec6

  Copy full SHA for 7b5fec6

• perf: keep only top tile value row at 16k — 22.87 → 24.32 tok/s (+6.3%)

  

  Autoresearch Agent

  committed
  f4317c0

  Copy full SHA for f4317c0

• perf: fix 16k sparse score scaling — 22.33 → 22.87 tok/s (+2.4%)

  

  Autoresearch Agent

  committed
  eff8c12

  Copy full SHA for eff8c12

• perf: use packed kv prompt flash — 1960.9 → 1999.6 tok/s (+2.0%)

  

  Autoresearch Agent

  committed
  355e9bf

  Copy full SHA for 355e9bf

• perf: half-output prompt qkv path — 1926.9 → 1960.9 tok/s (+1.8%)

  

  Autoresearch Agent

  committed
  46672b1

  Copy full SHA for 46672b1

• perf: split packed long-KV decode — 60.3 → 150.9 tok/s (+150.3%)

  

  Autoresearch Agent

  committed
  e815937

  Copy full SHA for e815937

• perf: word-align turboquant base decode — 16.47 → 17.33 tok/s (+5.2%)

  

  Autoresearch Agent

  committed
  f1e616f

  Copy full SHA for f1e616f

• perf: use fp16 K/V for prompt flash attention — 1748.3 → 1910.0 tok/s (+9.2%)

  

  Autoresearch Agent

  committed
  9fe0c7a

  Copy full SHA for 9fe0c7a

• perf: use MPS for packed prompt projections — 1356.0 → 1748.3 tok/s (+28.9%)

  

  Autoresearch Agent

  committed
  63e6d20

  Copy full SHA for 63e6d20

• perf: bitonic turboquant outlier selection — 13.25 → 16.47 tok/s (+24.3%)

  

  Autoresearch Agent

  committed
  ab6d504

  Copy full SHA for ab6d504

• perf: use MPS for packed prompt QKV — 1159.0 → 1356.0 tok/s (+17.0%)

  

  Autoresearch Agent

  committed
  fd2291a

  Copy full SHA for fd2291a

• perf: refine turboquant decode composition benchmarks

  

  Autoresearch Agent

  committed
  8264370

  Copy full SHA for 8264370

• perf: add exact turboquant decode breakdown harness

  

  Autoresearch Agent

  committed
  4f428b9

  Copy full SHA for 4f428b9

• perf: restore fused turboquant decode qkv — 12.50 -> 13.25 tok/s (+6.0%)

  

  Autoresearch Agent

  committed
  6a9a471

  Copy full SHA for 6a9a471

• perf: remove redundant turboquant decode barrier — 12.33 -> 12.50 tok/s (+1.4%)

  

  Autoresearch Agent

  committed
  b6f10ed

  Copy full SHA for b6f10ed

• perf: lazy aggressive turboquant rescale — 11.62 -> 12.33 tok/s (+6.1%)

  

  Autoresearch Agent

  committed
  8bd3cf1

  Copy full SHA for 8bd3cf1

• perf: flash prompt attention — 526.8 → 1169.6 tok/s (+122.0%)

  

  Autoresearch Agent

  committed
  88ccdbd

  Copy full SHA for 88ccdbd

Commits on Mar 28, 2026

• perf: fuse packed decode cache writes — 59.5 → 60.3 tok/s (+1.4%)

  

  Autoresearch Agent

  committed
  8a57a17

  Copy full SHA for 8a57a17

• perf: pack long-kv decode cache — 42.2 → 59.5 tok/s (+40.8%)

  

  Autoresearch Agent

  committed
  743d7ae

  Copy full SHA for 743d7ae

• perf: turboquant tiled decode breakdown — 11.21 -> 12.06 tok/s (+7.6%)

  

  Autoresearch Agent

  committed
  c5596de

  Copy full SHA for c5596de

• perf: vectorize packed prefill matmul — 503.3 → 558.6 tok/s (+11.0%)

  

  Autoresearch Agent

  committed
  745d704

  Copy full SHA for 745d704

• refactor: add matrix prefill projection helpers

  

  Autoresearch Agent

  committed
  de785a1

  Copy full SHA for de785a1

• refactor: extract shared prefill prerequisites for rewrite

  

  Autoresearch Agent

  committed
  381dd95

  Copy full SHA for 381dd95