jgrusewski
|
7cb8c305df
|
docs: GPU optimization full sweep design — training binary, PPO parity, NCCL multi-GPU
Audit found training binary bypasses all DQNTrainer/PpoTrainer infrastructure
and explicitly disables Rainbow DQN. Design covers 6 sections:
- P0: Wire train_baseline_rl.rs to Trainers
- P1: PPO mixed precision + gradient accumulation + Liquid/TLOB detach fixes
- P2: Mamba2 HardwareBudget + CUDA pipeline wiring + NCCL multi-GPU
- P3: Ensemble parallel inference + BF16 benchmarks
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-03-01 21:25:36 +01:00 |
|