13bf277cd6
Implements §4.2/4.3/4.4 + §5 of Phase 4 spec.
Three new CUDA kernels:
1. rl_dueling_q_bellman_target.cu — argmax over target composed_Q at
s_{t+1}, build target_value = r + γ^n × (1-done) × max_Q. Grid
(B,1,1), block (N_ACTIONS,1,1). Reads γ via per-sample n_step_gammas
passed by trainer (matches PER convention).
2. rl_dueling_q_loss_and_grad.cu — scalar Huber loss on
(target − online_composed_Q[taken]). Emits per-batch loss and
grad_composed (only taken action has nonzero gradient — this is
single-Q scalar regression, not distributional CE). Grid (B,1,1),
block (N_ACTIONS,1,1). Threads write zero into non-taken cells.
3. rl_dueling_q_decompose_and_bwd.cu — decompose grad_composed →
grad_V + grad_A via mean-subtraction Jacobian:
grad_V[b] = Σ_a grad_composed[b, a] = grad_composed[b, a_taken]
grad_A[b, a] = grad_composed[b, a] − (1/N) × grad_V[b]
Then computes per-batch weight gradients via outer product with h_t:
grad_w_v_pb[b, c] = grad_V[b] × h_t[b, c]
grad_b_v_pb[b] = grad_V[b]
grad_w_a_pb[b, c, a] = grad_A[b, a] × h_t[b, c]
grad_b_a_pb[b, a] = grad_A[b, a]
Grid (B,1,1), block (HIDDEN_DIM,1,1). Thread 0 also writes biases.
DuelingQHead Rust API (3 new methods):
- build_bellman_target(target_composed_q, r, dones, n_step_gammas, B, out)
- compute_loss_and_grad(online_composed_q, target_value, actions, B,
loss_pb, grad_composed_out)
- decompose_and_backward_to_weights(h_t, grad_composed, actions, B,
grad_w_v_pb, grad_b_v_pb,
grad_w_a_pb, grad_b_a_pb)
Caller (trainer) is responsible for reduce_axis0 of per-batch grads
→ final weight gradients [HIDDEN_DIM], [1], [HIDDEN_DIM × N_ACTIONS],
[N_ACTIONS]. Will be wired in Phase 4.2.
Soft-update of target weights still TODO — recommend reusing
dqn_target_soft_update kernel pattern in Phase 4.2.
Validated:
- cargo build --release clean
- integrated_trainer_smoke (1 step) passes
Per pearl_complement_internal_loss_vs_external_consumers: this loss
path is fully internal to DuelingQHead. No downstream consumer ever
sees composed_Q or grad_composed. The four prior session failure
modes are structurally impossible.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <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.