foxhunt

Files

jgrusewski 68e83bef0e fix: async adam_step HtoD + batch_size=0 auto — root cause of >128 hang

The batch_size >128 hang was caused by cudarc's synchronous
memcpy_htod for the adam_step counter. At batch_size=128 the
sync completes fast enough, but at 509+ the pipeline drain
from the sync interacts with CUDA Graph replay timing and
deadlocks the stream.

Replaced all 3 memcpy_htod(&[self.adam_step]) calls with raw
cuMemcpyHtoDAsync_v2 — zero pipeline drain, zero CPU sync.

Production TOML set to batch_size=0 (AutoBatchSizer drives it).
AutoBatchSizer caps at 8192 for RL training.

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

2026-04-02 08:19:56 +02:00

dqn-hyperopt.toml

feat(generalization): CQL alpha 0.1→1.0, gradient budget 15%→25%

2026-03-29 23:40:30 +02:00

dqn-localdev.toml

feat: MBP-10 → imbalance bars as explicit data source choice

2026-03-31 11:08:02 +02:00

dqn-production.toml

fix: async adam_step HtoD + batch_size=0 auto — root cause of >128 hang

2026-04-02 08:19:56 +02:00

dqn-smoketest.toml

fix: test reads threshold from dqn-smoketest.toml — no hardcoded values