Files
foxhunt/crates
jgrusewski 30f3156369 diag: RECAPTURE_DIAG — per-buffer norms after graph recapture
Fires for 3 steps after any CUDA graph recapture. Reports:
- d_val/adv_norm: blended d_logits (C51 × alpha + MSE × (1-alpha))
- mse_val/adv_norm: MSE-only scratch buffers
- grad_norm: backward output grad_buf

Purpose: diagnose C51 grad_norm=0 on H100 at batch=16384 (epoch 2+).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-10 08:34:01 +02:00
..