Repository navigation

perf(ArcInfer/ArcMoE): close the F32 seam — 5 sites, ~950 launches + 258 H2D copies per decode forward - #188

Merged

heydryft merged 27 commits into

masterfrom

agent/f32-cast-fuse

Aug 20, 2026

Merged

perf(ArcInfer/ArcMoE): close the F32 seam — 5 sites, ~950 launches + 258 H2D copies per decode forward#188
heydryft merged 27 commits into
masterfrom
agent/f32-cast-fuse

Commits on Aug 20, 2026
- feat(ArcKV/Fp8): fused E4M3 quantize+dequantize CUDA kernel, default path
  
  heydryft
  and
  claude
  committed
  66e0732
- fix(ArcKV/Fp8): inline-PTX IEEE ops (fast-math rewrote __f*_rn to .ftz) + box paths
  
  heydryft
  and
  claude
  committed
  e10cdea
- test(ArcKV/Fp8): exhaustive 2^32 E4M3 sweep + its negative control
  
  heydryft
  and
  claude
  committed
  259369b
- test(ArcLab): per-step D2H + launch counter, validated on a known answer
  
  heydryft
  and
  claude
  committed
  112cbb4
- test(ArcKV/Fp8): nsys count harness with lock + VRAM gate
  
  heydryft
  and
  claude
  committed
  41a9343
- test(ArcLab): count CUDA API calls per step too, incl. the invisible sync
  
  heydryft
  and
  claude
  committed
  ad75bdc
- fix(ArcLab): hold the bench lock across GPU work ONLY, not report export
  
  heydryft
  and
  claude
  committed
  02c3894
- test(ArcKV/Fp8): interleaved A/B driver with per-leg lock + Xid check
  
  heydryft
  and
  claude
  committed
  e57b959
- measure(ArcKV/Fp8): 44 -> 1 D2H/step; fused kernel 1.98 us/call [MEASURED]
  
  heydryft
  and
  claude
  committed
  488e458
- fix(ArcKV/Fp8): ARC_KV_FP8_MODE was a lying switch — rename, and stop typos selecting GpuApprox
  heydryft
  committed
  46c99e1
- fix(build): EXPECTED_KERNEL_COUNT 40 -> 41 — this branch adds arc_kvquant.cu
  heydryft
  committed
  3a71d03
- perf(arckv): make candle's caching allocator reachable and on for decode
  
  heydryft
  and
  claude
  committed
  dd5583c
- perf(arckv): bound the caching allocator, and print the counters that prove it
  
  heydryft
  and
  claude
  committed
  7d91c11
- fix(build): report_alloc_cache_step takes a reference
  
  heydryft
  and
  claude
  committed
  f9d496a
- chore(deps): candle 859c49c8 — allocator test fixes
  
  heydryft
  and
  claude
  committed
  fa1d2ef
- chore(deps): candle 89ab14ef — no eviction inside the capture window
  
  heydryft
  and
  claude
  committed
  65dc0fb
- fix(ci): spell 'unparsable' the way the typos gate expects
  
  heydryft
  and
  claude
  committed
  e4eb59d
- fix(deps): regenerate Cargo.lock for the candle 89ab14ef pin
  heydryft
  committed
  d71fec0
- perf(ArcMoE): fuse the V4 router region — 18+9 launches become 2, and fix the sinkhorn local-memory spill
  
  heydryft
  and
  claude
  committed
  9f11090
- fix(ArcMoE): cuda-gated build — device_ptr guard lifetime and monomorphic wrap closure
  
  heydryft
  and
  claude
  committed
  a92cf6d
- test(ArcMoE): extend the build-wiring tripwire to hc_fused.cu
  
  heydryft
  and
  claude
  committed
  48ae6b7
- docs(ArcMoE): cite the measured H200 numbers, not the estimates
  
  heydryft
  and
  claude
  committed
  ab5fbd5
- perf(ArcInfer/ArcMoE): fuse the mHC epilogue and post-step — 14 launches per mixing point become 7
  
  authored and
  heydryft
  committed
  179e405
- perf(ArcInfer/ArcAttention): fuse the V4 Q/K pre-attention block -- 16 launches per layer become 1
  
  authored and
  heydryft
  committed
  cd647ce
- perf(ArcInfer/ArcMoE): close the F32 seam — 5 sites, ~950 launches + 258 H2D copies per decode forward
  
  heydryft
  and
  claude
  committed
  1f6ef9d
- test(ArcGate): per-site A/B accounting, and name the two comparisons that are measurements
  
  heydryft
  and
  claude
  committed
  671a0c9
- docs(ArcInfer/ArcMoE): record the measured F32-seam result
  
  heydryft
  and
  claude
  committed
  b39a206

Commits

Commits on Aug 20, 2026