875f263ec9
Backward pass dX computation now uses f32 scratch buffers with bf16 staging for the backward chain. Pattern: f32 GemmEx → relu_mask (f32) → f32→bf16 cast → next layer reads bf16 dY. Changes: - 6 bw_d_h_* scratch buffers: CudaSlice<half::bf16> → CudaSlice<f32> - New bw_dy_bf16_staging: shared bf16 buffer for layer transitions - backward_fc_layer dX: gemmex_bf16 → gemmex_bf16_acc_f32 (f32 output) - launch_dx_only: gemmex_bf16 → gemmex_bf16_acc_f32 (f32 with beta) - relu_mask_kernel: reads/writes f32 (no bf16 clamp needed) - f32_to_bf16_cast_kernel: ±500 clamp at type boundary (in backward_kernels.cu) - cast_dx_to_staging: f32 scratch → bf16 staging per layer - IQN/ensemble backward: bf16→f32 cast for dX input, f32→bf16 for dY output - bw_d_h_s2_as_bf16(): attention backward receives bf16 via staging Hyperparameters updated for f32 Adam: - learning_rate: 1e-5 → 1e-4 (updates must exceed bf16 shadow step ~1e-3) - adam_epsilon: 1e-3 → 1e-8 (standard Adam, bf16 workaround no longer needed) - grad_norm NaN skip kept as defense-in-depth (source still under investigation) 895/895 unit + 359/359 ml-dqn tests pass. 7-11/11 smoke tests (intermittent NaN from unknown source — NOT backward dX). Co-Authored-By: Claude Opus 4.6 (1M context) <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;