789fb50dfe
Add cp.async.bulk.shared::cta.global tile loads guarded by #if __CUDA_ARCH__ >= 900 in common_device_functions.cuh: - cooperative_load_tile_tma(): thread-0-only bulk copy via inline PTX, freeing 31 warp threads for compute overlap. 16KB chunks with commit_group/wait_group barrier. - cooperative_load_tile_float4(): renamed original for fallback. - cooperative_load_tile(): dispatch wrapper (compile-time selection). Architecture-aware NVRTC compilation (compile_ptx_for_device): - Queries GPU compute capability, passes -arch=compute_XX to NVRTC. - Enables __CUDA_ARCH__ in kernels so TMA guard activates on Hopper. - Wired into all 3 runtime compilation sites (experience collector, backtest evaluator, PPO collector). Also fixes pre-existing clippy: vh * 1 identity op in weight estimate. 79 cuda_pipeline tests pass, 0 clippy errors. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ml
10-model ML ensemble for the Foxhunt HFT system, built on Candle v0.9.1.
Models
- DQN (Rainbow) — deep Q-network with prioritized replay, dueling heads, noisy nets
- PPO — proximal policy optimization with GAE, LSTM policies, clip-higher
- TFT — temporal fusion transformer for multi-horizon forecasting
- Mamba2 — state space model for sequence prediction
- Liquid Networks — biologically inspired networks for non-stationary data
- TLOB — transformer-based limit order book analysis
- KAN — Kolmogorov-Arnold networks
- xLSTM — extended LSTM architecture
- TGGN — temporal graph neural network
- Diffusion — diffusion-based generative model
Key Modules
ensemble— model ensemble coordination and confidence aggregationhyperopt— PSO-based hyperparameter optimization with per-model adapterstrainers— unified training loops (DQN, PPO, supervised)inference—InferenceAdaptertrait for predictioncheckpoint— model checkpointing and restorationevaluation— walk-forward evaluation pipeline
Usage
use ml::dqn::DQN;
use ml::ppo::PpoTrainer;