Repository navigation

The 794 ms decode step, itemised: lm_head's 244x, the duplicate K/V, and the 91 ms named - #202

Merged

heydryft merged 8 commits into

agent/decode-remainder-200ms

Aug 21, 2026

Merged

The 794 ms decode step, itemised: lm_head's 244x, the duplicate K/V, and the 91 ms named#202
heydryft merged 8 commits into
masterfrom
agent/decode-remainder-200ms

Commits on Aug 21, 2026
- fix(ArcGate): the citation gate had never opened a .rs file — 310 citations ungated
  
  authored and
  heydryft
  committed
  0595ba0
- perf(ArcKernels): lm_head re-read its 1.06 GB weight 256 times — flatten rank-3 in UnquantLinear
  
  authored and
  heydryft
  committed
  a187039
- fix(ArcGate): the rental gate certified "V4 fits in 60 GB" from a run that loaded on CPU
  
  authored and
  heydryft
  committed
  10448bb
- docs(ArcGate): commit the seven mission docs the code cites but nobody else could read
  
  authored and
  heydryft
  committed
  2fde72a
- chore(ArcGate): waive the grep pattern `sampl` quoted in BUDGET_V4_B1.md
  
  authored and
  heydryft
  committed
  c1c02a4
- perf(ArcInfer): the decode cohort crossed to the GPU one sequence at a time
  
  authored and
  heydryft
  committed
  1356f95
- fix(ArcKernels): exclude CPU from the rank-3 flatten — it is the oracle, and it moved
  
  authored and
  heydryft
  committed
  2fb4882
- chore(ArcKernels): drop a needless `Ok(..?)` in the new unquantized test helper
  
  authored and
  heydryft
  committed
  5e12594