jgrusewski dac9cfef5c spec: bypass cudarc for hot path — raw CUDA driver API + mega-kernel fusion ...

cudarc adds 18.8ms/step overhead at b=256 (GPU kernels take 0.16ms).
Phase 1: raw cuLaunchKernel wrapper. Phase 2: pre-extracted raw ptrs.
Phase 3: fused mega-kernels (20 launches → 3).
Target: 1ms/step → 500+ sps.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-26 09:39:17 +02:00

..

infra

…

plans

design(foxhuntq): architectural pivot away from per-bar DQN to decoupled Belief Bus + Conformal DRL

2026-05-14 19:21:25 +02:00

superpowers

spec: bypass cudarc for hot path — raw CUDA driver API + mega-kernel fusion

2026-05-26 09:39:17 +02:00

dqn-backward-nan-audit.md

…

dqn-gpu-hot-path-audit.md

…

dqn-named-dims.md

…

dqn-wire-up-audit.md

fix(ml): migrate test uploads from clone_htod to mapped-pinned

2026-05-22 20:39:32 +02:00

h_s2_consumers_audit.md

…

health_diag_inventory.md

…

isv-slots.md

feat(regime): vol_ref floor controller + cross-invocation disk persist

2026-05-16 09:22:20 +02:00

lookahead-bias-audit-2026-04-28.md

fix(audit): lookahead housekeeping — purge gap + per-fold NormStats

2026-05-09 13:52:00 +02:00

ml-supervised-to-dqn-concept-audit.md

…

sp18-wireup-audit.md

feat(loader): instrument_id filter + outdated test fix + sp18 fingerprint

2026-05-22 15:56:31 +02:00