Repository navigation

perf(ArcAttention): V4 RoPE was launched per-sequence — 99,072 launches/step -> 301 - #198

Merged

heydryft merged 6 commits into

release/openrouter-readyfrom

perf/rope-cohort-uniform-offsets-194

Aug 21, 2026

Merged

perf(ArcAttention): V4 RoPE was launched per-sequence — 99,072 launches/step -> 301#198
heydryft merged 6 commits into
release/openrouter-readyfrom
perf/rope-cohort-uniform-offsets-194

Commits on Aug 21, 2026
- perf(ArcAttention): batch V4 RoPE over the cohort — 99,072 launches/step -> 301
  
  authored and
  heydryft
  committed
  1e6cff1
- perf(ArcAttention): un-deaden the fused qk_norm_rope kernel above batch 1
  
  authored and
  heydryft
  committed
  44f3c52
- fix(ArcAttention): one cohort predicate, one counter pair, per-thread test counts
  
  heydryft
  and
  claude
  committed
  e6d2856
- fix(ArcAttention): the batched fused qk_norm_rope path is OPT-IN
  
  heydryft
  and
  claude
  committed
  eea014c
- perf(ArcAttention): gather the RAGGED cohort's RoPE — one rank-3 rope_i, opt-in ARC_ROPE_COHORT
  
  heydryft
  and
  claude
  committed
  192f8ff
- fix(ArcAttention): bounds-guard the cohort gather — both ragged arms must fail LOUD
  
  heydryft
  and
  claude
  committed
  227c34f