Files
foxhunt/config
jgrusewski 5d78108c77 feat(bf16): BF16-tuned smoke test config + huber_delta TOML support
dqn-smoketest.toml: lr 3e-5→1e-5, gradient_clip 10→1,
  huber_delta 10→1, q_clip ±200→±50, reward_scale 10→1,
  noisy_sigma 0.5→0.3, spectral_norm 3→1.5

training_profile.rs: add huber_delta to TrainingSection + apply_to

Smoke test runs 3 epochs but Q-values diverge — C51 forward produces
overflow logits in bf16. Needs CUDA graph buffer size audit (graph
captured with F32 byte counts may have wrong bf16 sizes on replay).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-28 15:57:17 +01:00
..