7349713ca8
Output layer GemmEx now writes f32 (CUDA_R_32F C-matrix) instead of bf16. When the f32 accumulated dot product exceeded bf16 max (~65504), the bf16 C-matrix write produced Inf/NaN → propagated through softmax → NaN loss. With f32 output, no truncation overflow is possible. Forward pass changes: - 6 logit buffers: CudaSlice<half::bf16> → CudaSlice<f32> (on/tg/on_next × v_logits/b_logits) - gemmex_bf16_to_f32(): BF16 A/B, F32 C — same tensor core throughput - launch_add_bias_f32_raw(): uses add_bias_f32_kernel (f32 in, f32 out) - Loss kernels: 12 logit params changed to const float* (no bf16→float cast) - expected_q_kernel: full f32 rewrite (was bf16 arithmetic) - cql_grad_kernel: logit inputs as float*, internal computation f32 - ensemble_kernels: logit inputs as float*, softmax/KL in f32 - gpu_experience_collector: exp_v_logits/exp_b_logits → CudaSlice<f32> - gpu_backtest_evaluator: chunked logit buffers → CudaSlice<f32> - ensemble_logits_buf in fused_training: CudaSlice<f32> Per-sample NaN guard kept as safety net for bf16 reward/done/IS-weight edge cases (not the primary fix — f32 logits are the root cause fix). Hidden layers stay bf16 — ReLU bounds them. Gradient buffers stay bf16/f32 (already converted). No tensor core throughput loss (same CUBLAS_COMPUTE_32F). 50-epoch convergence: all 50 epochs complete, loss=4.12, Q=1.33, grad=0.49. 895/895 unit tests, 9/9 smoke tests pass. 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.