Files
foxhunt/crates
jgrusewski 36d076b344 feat(iql): full pipeline wired — dual IQL, per-sample support, branch scales, gap exploration
- IQL mandatory: Option<GpuIqlTrainer> → GpuIqlTrainer (high + low tau)
- Full pipeline in submit_aux_ops: populate_q_out → gather → train both
  → advantages → adv_sigma → per_sample_support → branch_scales → gap → epsilon
- Delete v_range wrappers from FusedTrainingCtx
- Delete adapt_v_range calls from training_loop.rs
- Rename v_range_ptr → per_sample_support_ptr in experience collector
- Fixed eval v_range in metrics.rs (wide fixed range for argmax)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-13 15:26:39 +02:00
..