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

..

audit

audit: 21 CUfunction conflicts — vaccine uses 15+ forward_child kernels ungraphed

2026-04-18 21:45:27 +02:00

cleanup

cleanup: update scoreboard — iter 10 (PINMEM-007/013 rename + BORROW-001/ROMEM-003 u32 fix)

2026-04-21 00:13:32 +02:00

memos

memo(crt-a): forward_only cost investigation — Case 2 (stateless K-window)

2026-05-20 17:43:52 +02:00

notes

baseline(per-horizon-cfc): 3-seed h6000 baseline JSON skeleton + retrieval gap

2026-05-21 16:10:48 +02:00

plans

spec+plan: DQN concepts adoption — PopArt, spectral norm, outcome head, enrichment E1-E8

2026-05-26 08:59:51 +02:00

results

docs(claude): foxhunt agents & skills rollout — phase 1 close-out

2026-05-10 22:19:28 +02:00

specs

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

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