Files
foxhunt/crates
jgrusewski d572dc466a perf(cuda): batch DQN metrics readbacks — 15 DtoH transfers → 4
metrics.rs: 4 fixes batching individual to_scalar() GPU syncs:
- Q-value stats: 4 transfers → 1 (cat [min,max,mean,var])
- Q diagnostics: 8 transfers → 1 (cat [gaps + per-action avgs])
- get_q_values: N transfers → 1 (bulk to_vec1)
- Validation Sharpe: 2 transfers → 1 (cat [mean,var])

Zero per-step DtoH leaks in training_loop.rs or metrics.rs.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-18 08:29:26 +01:00
..