jgrusewski d070dccf6e spec: kernel optimization — cuBLAS projections + batch-parallel scans ...

nsys data: custom kernels = 95% of GPU time. mamba2_scan_backward = 75%.
Root cause: one-thread-per-weight anti-pattern serializes B=8192 samples.
Fix: cuBLAS for projections, lightweight scans for sequential state.
Target: 1950ms/step → <50ms/step.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-19 12:10:49 +02:00

..

infra

feat(infra): GPU training launcher with local/cloud routing

2026-02-23 11:19:58 +01:00

plans

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

superpowers

spec: kernel optimization — cuBLAS projections + batch-parallel scans

2026-04-19 12:10:49 +02:00