foxhunt

Files

jgrusewski 632c780c00 feat(ml): GPU max performance phase 2 — C51 re-enabled, buffer pooling, kernel sizing

Phase 2 GPU optimizations for L4→H100 scaling:

- Re-enable C51 distributional RL (BUG #36 scatter_add gradient flow VERIFIED)
- Add GpuBufferPool for zero-alloc walk-forward fold transitions
- Add DoubleBufferedLoader for CUDA stream overlap during fold swaps
- Add EpochPrefetcher for background data preparation (overlaps I/O with GPU)
- Add optimal_launch_dims() for dynamic CUDA kernel block/grid sizing (32→256)
- Wire real INT8 quantization into QuantizedTFT via quantize_varmap_parallel()
- Wire DoubleBufferedLoader + EpochPrefetcher into DQN trainer
- Wire optimal_launch_dims into DQN + PPO GPU experience collectors
- Fix flaky hot_swap latency test (100μs→2ms threshold for debug builds)
- Fix missing mixed_precision field in trading_service PPOConfig

Full Rainbow DQN now enabled by default (all 6 components + IQN + CQL).
2,437 ml tests pass, 0 failures. Workspace compiles clean.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-03-01 21:25:36 +01:00

backtesting

perf: eliminate heap-alloc Decimal→float casts across 19 files (36 instances)

2026-02-28 02:29:04 +01:00

common

refactor: split common/types.rs (4909 LOC) into 10 focused modules

2026-03-01 20:32:52 +01:00

config

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

ctrader-openapi

refactor: restructure repo — crates/, bin/, testing/ layout