Commits

Branch selector

 main

User selector

All users

All time

Commit History

Commits on Jun 3, 2026

• Merge remote-tracking branches 'origin/feat/perf-01-prefill-tensorcore', 'origin/feat/perf-02-decode-attention-splitkv' and 'origin/feat/perf-04-moe-decode-occupancy' into feat/perf-all-13-14-16

  cchuter

  committed
  0b6035b

  Copy full SHA for 0b6035b

Commits on Jun 2, 2026

• perf(mgpu): fill SMs for routed-MoE gate/up decode via finer launch geometry

  Show description for bc0ce8e
  

  cchuter

  and

  claude

  committed
  bc0ce8e

  Copy full SHA for bc0ce8e

• fix(mgpu): apply exact window logic when sizing split-KV S

  Show description for 58b5053
  

  cchuter

  and

  claude

  committed
  58b5053

  Copy full SHA for 58b5053

• perf(mgpu): add opt-in split-KV/flash-decode attention kernels

  Show description for cd51fe5
  

  cchuter

  and

  claude

  committed
  cd51fe5

  Copy full SHA for cd51fe5

• perf(cuda): safe smaller q8->f16 cache reserve on high-VRAM cards

  Show description for 35f913e
  

  cchuter

  and

  claude

  committed
  35f913e

  Copy full SHA for 35f913e

• test(mgpu): bump test_engine_correctness multi-tier budgets to clear pre-subtract

  Show description for ef64b69
  

  cchuter

  and

  claude

  committed
  ef64b69

  Copy full SHA for ef64b69

• fix(mgpu): plumb ctx into placement, per-layer KV math, auto-mode reserve, upfront refusal (#12)

  Show description for 5f94b24
  

  cchuter

  and

  claude

  authored
  5f94b24

  Copy full SHA for 5f94b24

Commits on May 28, 2026

• chore: merge upstream/main into fork — sync v0 multi-GPU work with upstream improvements (#11)

  Show description for 4817ce7
  

  12 peopleauthored
  4817ce7

  Copy full SHA for 4817ce7

• chore: gitignore docs/ entirely and untrack existing docs (#10)

  Show description for 5cc3c74
  cchuter

  authored
  5cc3c74

  Copy full SHA for 5cc3c74

• test(mgpu): correctness checks — smoke output + deterministic token compare (#9)

  Show description for 7d6a83b
  

  cchuter

  and

  claude

  authored
  7d6a83b

  Copy full SHA for 7d6a83b

• feat(mgpu): wave 3 v0 baseline — bench harness + report (#8)

  Show description for ebd073d
  

  cchuter

  and

  claude

  authored
  ebd073d

  Copy full SHA for ebd073d

• feat(mgpu): wave 2 — wire --gpu-vram and --gpu-devices into all four binaries (#7)

  Show description for d1a9ff9
  

  cchuter

  and

  claude

  authored
  d1a9ff9

  Copy full SHA for d1a9ff9

• feat(mgpu): wave 3a Half-B — engine-side multi-tier dispatch (#6)

  Show description for afedc61
  

  cchuter

  and

  claude

  authored
  afedc61

  Copy full SHA for afedc61

• feat(mgpu): wave 3a Half-A — CUDA-side per-device plumbing (#5)

  Show description for 618644a
  

  cchuter

  and

  claude

  authored
  618644a

  Copy full SHA for 618644a

Commits on May 27, 2026

• feat(mgpu): engine + placement scaffolding (wave 2) (#4)

  Show description for 68bd41a
  

  cchuter

  and

  claude

  authored
  68bd41a

  Copy full SHA for 68bd41a

• feat(mgpu): per-device selective model cache mechanism (wave 1) (#3)

  Show description for a427e07
  

  cchuter

  and

  claude

  authored
  a427e07

  Copy full SHA for a427e07

• feat(mgpu): device-aware CUDA plumbing (wave 1) (#2)

  Show description for 78a29c0
  

  cchuter

  and

  claude

  authored
  78a29c0

  Copy full SHA for 78a29c0

• feat(mgpu): monotonic-contiguous layer-placement packer (wave 1) (#1)

  Show description for 041083f
  

  cchuter

  and

  claude

  authored
  041083f

  Copy full SHA for 041083f

• docs: address rounds 3-4 of codex spec review

  Show description for 5dee40f
  

  cchuter

  and

  claude

  committed
  5dee40f

  Copy full SHA for 5dee40f

• docs: multi-GPU pipeline-parallel v0 design

  Show description for abf729e
  

  cchuter

  and

  claude

  committed
  abf729e

  Copy full SHA for abf729e

• chore: ignore claude-dev-team workspace and symlinks

  Show description for 596e554
  

  cchuter

  and

  claude

  committed
  596e554

  Copy full SHA for 596e554

Commits on May 21, 2026

• docs: remove local q8 planning notes

  cchuter

  committed
  7780293

  Copy full SHA for 7780293

• docs: record q8_k branch handoff

  cchuter

  committed
  ce376a7

  Copy full SHA for ce376a7

• docs: record q8_k routed expert verification

  cchuter

  committed
  8470cb6

  Copy full SHA for 8470cb6

• test: cover multi-block q8_k metal moe

  cchuter

  committed
  d6f416d

  Copy full SHA for d6f416d

• test: cover q8_k metal routed moe

  cchuter

  committed
  9a151ae

  Copy full SHA for 9a151ae

• fix: compute full q8_k metal matvec rows

  cchuter

  committed
  6e23616

  Copy full SHA for 6e23616

• feat: route q8_k experts on metal

  cchuter

  committed
  0bb142b

  Copy full SHA for 0bb142b

• fix: harden q8_k routed cpu scratch

  cchuter

  committed
  2189ecb

  Copy full SHA for 2189ecb

• feat: route q8_k experts on cpu

  cchuter

  committed
  9bac3f4

  Copy full SHA for 9bac3f4

• feat: add cpu q8_k routed expert kernels

  cchuter

  committed
  6ee78c0

  Copy full SHA for 6ee78c0

• docs: clarify q8_k quantizer docs

  cchuter

  committed
  4fafb68

  Copy full SHA for 4fafb68

• feat: emit q8_k routed expert tensors

  cchuter

  committed
  c40a4f2

  Copy full SHA for c40a4f2

• docs: plan q8_k routed expert support

  cchuter

  committed
  8809bab

  Copy full SHA for 8809bab

• docs: document q8_0 routed expert model

  cchuter

  committed
  faa1827

  Copy full SHA for faa1827