Files
foxhunt/docs
jgrusewski 9843de5e3d feat(sp14): B.7 — trainer struct fields for EGF kernels + concat scratch
Adds 4 CudaFunction handles + 1 scratch buffer to GpuDqnTrainer:
- sp14_q_disagreement_update_kernel
- sp14_alpha_grad_compute_kernel
- sp14_gradient_hack_detect_kernel
- sp14_dir_concat_qaux_kernel
- sp14_dir_qaux_concat_scratch: CudaSlice<f32> [B * (SH2 + 1)]

All loaded from precompiled cubins in trainer construction, mirroring
the SP13 aux_pred_to_isv_tanh / aux_sign_label kernel-loading pattern.
Per feedback_no_partial_refactor: handles + scratch are held but not
yet wired in. Subsequent tasks (B.9+) launch them.

Build: cargo check --workspace clean.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-05 19:41:12 +02:00
..