Files
foxhunt/crates
jgrusewski 937aa45ba9 fix(bf16): staging pipeline + upload_batch_gpu BF16 boundaries
- upload_batch: exclude actions from bf16 staging, upload i32 separately
- upload_batch_gpu: replace broken f32 DtoD with bf16 DtoD + proper
  actions bf16→i32 conversion via host roundtrip (64 elements, negligible)
- Remove dead mixed-precision conversion paths (bf16_states_buf etc.)

Smoke test now reaches CUDA graph capture + forward + loss + backward.
Next blocker: dqn_grad_norm_kernel launch args (shared memory or
buffer size from bf16 conversion).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-28 15:28:43 +01:00
..