8329ca4187
The bf16 backward chain destroyed gradient precision at batch=16384 with mean-reduced gradients (~6e-5). bf16's 8-bit mantissa couldn't represent these values, producing zero weight gradients on H100. This commit removes bf16 from the ENTIRE training pipeline: Forward pass: - cublasSgemm with CUBLAS_TF32_TENSOR_OP_MATH math mode (auto TF32 on Ampere+) - f32 master weights used directly (no bf16 shadow for forward) - All activation saves (h_s1, h_s2, h_v, h_b[0..3]) now f32 - States buffer f32 (pad_states_kernel outputs f32) - Bias kernels: pure f32 (removed 5 bf16 variants) Backward pass: - Single cublasSgemm GEMM (was 6 variants: bf16, bf16_acc_f32, f32dy, etc.) - f32 activations + f32 weights → no casts needed - relu_mask_kernel reads f32 activation (was bf16) - Removed: bf16 staging buffer, cast_dx_to_staging, all _f32dy duplicates Backtest evaluator: - All activation/state/weight buffers converted bf16→f32 - gather_states outputs f32 (kernel reads bf16 features, writes f32) - Weight flattening: bf16→f32 conversion via kernel Net: -1290 lines, +556 lines (734 lines removed) Rule: bf16 is ONLY for stored weight tensors (spectral norm). Everything else is f32. 19/19 smoke tests pass. Gradient norms healthy (0.39-1.03). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Foxhunt
Production HFT trading system in Rust.
Architecture
The workspace contains 32 crates organized as follows:
Core Libraries (16)
| Crate | Purpose |
|---|---|
trading_engine |
Order processing, FIX 4.4, IB TWS, SIMD, RDTSC timing |
risk |
VaR, Kelly, circuit breakers, kill switches, compliance |
risk-data |
Risk data types and shared structures |
trading-data |
Trading data types |
ml |
DQN Rainbow, PPO, TFT, Mamba2, ensemble inference |
ml-data |
ML data types and feature definitions |
data |
Market data ingestion and storage |
backtesting |
Replay engine, strategy tester |
adaptive-strategy |
Ensemble execution, microstructure analysis |
common |
Shared types, resilience, error handling |
storage |
S3 and local model storage |
model_loader |
Model serialization and loading |
market-data |
Market data feed handlers |
database |
PostgreSQL access layer (SQLx) |
config |
Configuration management |
tli |
CLI commands and tooling |
Services (8)
| Service | Purpose |
|---|---|
backtesting_service |
gRPC backtesting service |
broker_gateway_service |
FIX routing, broker connectivity |
trading_service |
Core trading operations |
ml_training_service |
Model training orchestration |
data_acquisition_service |
Market data acquisition |
trading_agent_service |
Autonomous trading agents |
api_gateway |
gRPC API gateway with auth |
web-gateway |
Axum REST + WebSocket gateway |
Frontend
web-dashboard/ -- React 19 + TypeScript + Vite + TradingView charts.
Building
# Check compilation (no PostgreSQL required)
SQLX_OFFLINE=true cargo check --workspace
# Run tests for a specific crate
SQLX_OFFLINE=true cargo test -p <crate> --lib
# Clippy
SQLX_OFFLINE=true cargo clippy --workspace
ML Models
Four production model architectures on Candle v0.9.1 with CUDA:
- DQN Rainbow -- Deep Q-Network with prioritized replay, dueling heads, noisy nets
- PPO -- Proximal Policy Optimization with GAE and LSTM policies
- TFT -- Temporal Fusion Transformer for multi-horizon forecasting
- Mamba2 -- State space model for sequence prediction
Each model has a standalone trainer and a UnifiedTrainable adapter for the hyperopt pipeline.
Infrastructure
- Git: Gitea at
git.fxhnt.ai(Tailscale-only), Scaleway DEV1-S - Observability: OpenTelemetry OTLP (env
OTEL_EXPORTER_OTLP_ENDPOINT) - Database: PostgreSQL with SQLx offline mode for CI
License
Proprietary. All rights reserved.