The GPU PER replay buffer had a hardcoded 4 GB MAX_BYTES limit that
rejected the auto-sizer's 10M-entry proposal on H100 (80 GB VRAM).
Now per_max_buffer_bytes() computes 20% of total VRAM (min 1 GB) and
flows through OptimalReplayConfig → DQNConfig → GpuReplayBufferConfig
so both subsystems agree on the budget.
Also fixes misleading regime detection log (indices 211/219 → 40/41)
and renames dqn_config_2025 → dqn_default_config.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Forward-port from worktree-gpu-hotpath-audit (2 commits):
1. Zero-roundtrip DQN experience collection via cuMemcpyDtoDAsync:
- GpuExperienceCollector uses device-to-device copies for state tensors
- Shared memory weight caching in GPU replay buffer
- NaN priority clamping on GPU (no CPU readback)
2. Eliminate all GPU→CPU roundtrips from training loop:
- GpuTrainResult: loss + grad_norm stay as GPU scalar tensors
- Single to_scalar() readback at epoch boundary (not per step)
- GPU-resident loss accumulation across training steps
- RegimeConditional head selection via GPU tensor ops
- Removed per-step NaN diagnostic checks (now epoch-level)
Also removes dead `states_tensor` field from ComputeLossResult and
fixes redundant field name clippy warning.
Expected: ~15-20% training throughput improvement on H100 by
eliminating synchronous GPU→CPU transfers in the inner loop.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three root causes fixed:
1. GPU PER + experience collector (cudarc) created dual CUDA allocator
fragmentation on GPUs ≤8 GB, making training impossible even at
batch_size=1. Added `use_gpu_replay_buffer` config flag; both GPU PER
and experience collector now disabled when VRAM ≤8192 MB.
2. Search space bounds were WIDENED instead of capped — max_batch_size
returning 4096 replaced the original 512 upper bound, and
max_hidden_dim_base_full returning 3072 replaced the original 1024.
Fixed with min() to only narrow, never widen.
3. VRAM estimator assumed GPU features always active, overcharging when
they're disabled on small GPUs. Now conditional: when
replay_buffer_capacity=0 (proxy for GPU PER disabled), collector/cudarc
costs are zero and fragmentation multiplier drops from 3× to 1.5×.
Additional small-GPU guard: GPUs ≤8 GB get clamped search space
(batch≤128, hidden≤512, atoms≤51, buffer≤50K) to fit 3 regime heads
+ C51 + noisy nets + dueling in limited VRAM.
Validated: 7/7 trials complete on RTX 3050 Ti 4 GB, zero OOM, best trial
Sharpe 9.8 with 50.6% win rate. Previous runs had 100% OOM failure rate.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Expand FeatureVector from 40 to 42 dimensions by including ADX(14) at
index 40 and CUSUM direction at index 41 from the existing CPU feature
extraction pipeline. This eliminates proxy-based regime classification
and enables GPU-native regime detection via tensor narrow/comparison ops.
Key changes:
- extraction.rs: wire RegimeADXFeatures + RegimeCUSUMFeatures into
extract_current_features_v2(), output 42 features per bar
- regime_conditional.rs: classify_regime_masks_gpu() creates per-regime
mask tensors entirely on GPU (ADX > 0.25 = trending, |CUSUM| > 0.7 =
volatile, else ranging). Zero CPU roundtrip in training hot path.
- trainer.rs/config.rs: state_dim 43→45 (no OFI), 51→53 (with OFI),
aligned dims unchanged (48/56). GPU batch insertion for all 3 heads.
- CUDA header: MARKET_DIM 40→42
- walk_forward.rs: FEATURE_DIM 40→42
- 42 files updated, all [f64;40]→[f64;42] propagated across workspace
Test results: ml=874/0, ml-dqn=354/0, ml-features=282/0, ml-core=274/0
Real data GPU smoke tests: 7/7 passed (OHLCV + OFI + trade enrichment)
Hyperopt baseline RL: 2 trials completed on local RTX 3050 Ti
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move TFT, Mamba-2, Liquid, TGGN, TLOB, KAN, xLSTM, and Diffusion model
implementations to ml-supervised. Bridge files (UnifiedTrainable adapters,
Checkpointable impls) stay in ml. Delete AsyncDataLoader (replaced by
StreamingDbnLoader + simple .chunks() batching). Remove empty ml-infra
scaffold — the remaining ml modules are too tightly coupled for clean
extraction, so ml stays as the orchestration facade.
- ml-supervised: 234 tests, 0 failures
- ml: 1687 tests, 0 failures
- Workspace: 0 compilation errors
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The GPU experience kernel had hardcoded STATE_DIM=54, MARKET_DIM=51,
NUM_ATOMS_MAX=51 but the host-side feature buffer uploads 40 features
per bar and networks use state_dim=56 with up to 100 atoms. Past ~702K
bars the stride mismatch caused out-of-bounds GPU reads → ILLEGAL_ADDRESS.
All dimension constants now use #ifndef guards so the NVRTC JIT compiler
receives the actual values via #define injection — zero runtime overhead.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Hyperopt varies hidden_dim_base (256–1024) but the CUDA experience
kernel hardcoded SHARED_H1=256, SHARED_H2=256. When hyperopt picked
larger dims, the kernel did matvec with wrong strides → illegal memory
access → training crash on first epoch.
Fix: wrap CUDA #defines in #ifndef guards and inject actual dims from
the dueling network config at NVRTC compile time. The kernel is now
specialized per-trial with the exact weight matrix dimensions — zero
runtime overhead, no dimension mismatch possible.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The OFI calculator's compute_ofi_from_file() never called feed_trade(),
leaving 3/8 features (VPIN, Kyle's Lambda, trade_imbalance) at zero in
production. Added compute_ofi_with_trades() that interleaves trades by
timestamp into the MBP-10 streaming callback. Updated DQN and PPO
hyperopt adapters to derive trades_dir as sibling of dbn_data_dir.
Smoke test validates: without trades VPIN=0/5000, with trades VPIN=5000/5000.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tests require a CUDA GPU and are #[ignore]d by default. Run with:
cargo test -p ml --test gpu_kernel_parity_test -- --ignored
Covers:
- Standard dueling forward: finite Q-values, valid action range [0,4]
- C51 distributional forward: Q-values bounded by atom support [-25,25]
- Candle vs kernel Q-value parity: argmax action consistency
- NoisyNet exploration: noise injection produces action diversity
- Weight extraction roundtrip: VarMap → CudaSlice → sync
- Distributional weight shapes: value_out [51,128], advantage_out [255,128]
- RMSNorm gamma extraction: initialized to 1.0
- Repeated kernel launches: 5 consecutive runs all produce finite output
All 8 tests pass on RTX 3050 Ti (4.49s total).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire 8 OFI features (OFI L1/L5, depth imbalance, VPIN, Kyle's lambda,
bid/ask slopes, trade imbalance) through the DQN training pipeline:
- Add mbp10_data_dir config field to DQNHyperparameters
- Dynamic state_dim: 43 (no OFI) or 51 (with OFI) based on config
- Compute OFI per bar during data loading, store on trainer
- Pass OFI features through regime_features slot in TradingState
- Configurable MBP-10 path with recursive .dbn/.dbn.zst discovery
- Add zstd auto-detection to DbnParser::parse_mbp10_file()
- Add --mbp10-data-dir CLI flag to train_baseline_rl
- Fix hardcoded [f64; 51] → FeatureVector51 ([f64; 40]) across
examples, walk_forward, GPU memory profile, and test fixtures
- Fix stale state_dim=51 in dqn_config_2025() and DQN tests
2747 tests pass, 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When use_noisy_nets=true (the conservative() default), epsilon never
decayed from 1.0 because (1) the trainer skipped update_epsilon() and
(2) DQNAgentType::set_epsilon() was a no-op for RegimeConditional agents.
This caused ALL training actions to be random — Q-values were learned but
never used for action selection.
Fix: set stored epsilon to 0.0 at training start when noisy nets are on.
Exploration is provided by NoisyLinear weight perturbation + the separate
noisy_epsilon_floor (5% safety floor for 45-action spaces).
Also fixes:
- Zstd-compressed .dbn file detection via magic bytes (0x28B52FFD)
- Test data path resolution using ancestors().find() for workspace root
- Test assertions updated for epsilon < 0.01 with noisy nets
2698 lib tests pass, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add live training metrics monitor CLI command (streaming & one-shot) using
the monitoring gRPC service. Update DQN tests to match post-fix defaults:
IQN disabled, CQL alpha=0.1, v_min/v_max widened, 26D search space.
- train.rs: `fxt train monitor [--once] [--model X] [--interval N]`
- Rewrite gradient collapse test for BF16 mixed precision awareness
- Update inference test config to match trainer defaults (IQN off, CQL on)
- Update production smoke test for 26D parameter space
- Add dqn_action_collapse_fix_test.rs verifying all 6 root cause fixes
- Add planning docs for monitoring service and epoch financial metrics
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Delete stub rainbow_agent.rs (fake select_action, hardcoded train loss)
and promote rainbow_agent_impl.rs to rainbow_agent.rs. Unify on the
real 8-field RainbowAgentMetrics from rainbow_config.rs, replacing the
4-field stub version (epsilon→exploration_rate, average_loss→current_loss).
5 files changed: -667/+447 lines, 2698+32 tests pass, 0 clippy.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
egobox is not a dependency (replaced by argmin PSO long ago).
The egobox_tuner.rs module was pure dead code — optimize_mamba2()
just returned a deprecation error. Deleted the module, its 26 tests,
and the integration test. Cleaned up stale egobox references in
ParameterSpace trait docs.
174 hyperopt tests pass, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
DQN: disable all exploration for production (warmup=0, epsilon=0,
noisy_nets=false, count_bonus=false), read feature_count from
checkpoint metadata instead of hardcoding 54, add loaded guard
and NaN check on predict output.
PPO: fix confidence formula — use act_with_log_prob() instead of
act() which returns value_estimate not log_prob, add loaded guard
and NaN check, remove WorkingPPO alias.
Liquid: add loaded flag for is_ready()/predict() guards.
Remove EgoboxOptimizer/EgoboxOptimizerBuilder backward-compat
aliases — replaced with canonical ArgminOptimizer everywhere.
323 trading_service tests, 200 hyperopt tests, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Integrate all PPO improvement modules into the core training paths:
- Symlog value predictions in compute_value_loss (MLP + LSTM)
- Adaptive entropy auto-tuning replaces fixed entropy_coeff
- Percentile P5/P95 advantage scaling for heavy-tailed returns
- DAPO clip_epsilon_high wired in all 7 PPOConfig construction sites
- Shape mismatch fix in adaptive_entropy (unsqueeze scalar)
2726 tests pass, 0 clippy errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
5 tests: training loop cycle, priority-weighted sampling, KS distribution
test, IS weight range validation, ring buffer overwrite correctness.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The real_ prefix was misleading — there is no fake data loader.
Mechanical rename across 18 source files, no logic changes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
DQN and PPO trainers now resolve hidden_dim_base from GPU VRAM when no
explicit override is given, so L4/L40S/H100 GPUs use proportionally
larger networks instead of being stuck at RTX 3050 Ti defaults (256).
- Add resolve_hidden_dim_base() tiered lookup (256/512/768/1024 by VRAM)
- Add network_param_count() for accurate model size estimation
- DQN: pre-compute hidden dims, use real param count for batch sizing
- PPO: add hidden_dim_base field, VRAM resolution in PpoTrainer::new()
- PPO hyperopt: raise hidden_dim_base ceiling from 2048 to 4096
- TFT hyperopt: expand hidden_sizes from [128,256,512] to 5 tiers
- Update stale estimates (DQN 50K→200K, PPO 100K→400K params)
- Fix pre-existing clippy lints in prefetch.rs and ppo.rs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Remove stale test reports, quick-start guides, benchmark analyses,
profiling reports, and tool artifacts from across the workspace.
Keeps only root README.md per crate/service.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Exposes gradient_accumulation_steps in PpoHyperparameters so hyperopt
and training binaries can configure effective batch scaling. The actual
accumulation logic already existed in PPO::update_mlp().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Quality audit found 2 dead code paths in the GPU optimization commit:
1. Data caching: preload_data() was defined on all 10 hyperopt adapters
but never called. Now wired in both hyperopt binaries (RL + supervised)
before the trial loop. Each model preloads training data once into
Arc<Vec<...>>, eliminating per-trial disk I/O.
2. PPO mixed precision: config.mixed_precision was stored but never used
in forward passes. Added forward_mixed() to PolicyNetwork and
ValueNetwork (same BF16/FP16 pattern as DQN's NetworkLayers). Stored
on network structs and auto-applied via forward(). Wired in
PPO::with_device() for MLP networks.
Also fixes missing mixed_precision field in 2 test files and
trading_service PPOConfig literal.
5 files changed, +152/-20. 2418 tests pass, workspace compiles clean.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 3 of the CUDA pipeline: both trainers now take a GPU-first path
for experience collection (128×500 = 64K experiences per kernel launch,
zero CPU-GPU roundtrips per timestep) with automatic CPU fallback.
- DQN: upload features alongside targets, GPU collection branch before
CPU loop, gpu_batch_to_experiences() conversion into replay buffer
- PPO: set_raw_market_data() for CudaSlice upload, GPU collection
branch bypasses collect_rollouts + prepare_training_batch entirely,
gpu_batch_to_trajectory_batch() conversion with in-kernel GAE
- Configurable GPU batch sizes (gpu_n_episodes, gpu_timesteps_per_episode)
and trading params (initial_capital, avg_spread) via hyperparameters
- SAFETY training diagnostics downgraded from warn! to debug!
- 4 new batch index-math validation tests in cuda_pipeline
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Split the build pipeline: one compile-services job builds all 8 service
binaries with PVC-backed sccache, saves as artifacts. Then 9 Kaniko jobs
just package pre-built binaries into slim runtime images (~30s each).
Before: 9 parallel Kaniko jobs each doing full cargo build --release
(~20min each, no sccache, 9x duplicated dep compilation)
After: 1 compile job with sccache (~5min cached) + 9 package jobs (~30s)
- Add compile stage between test and build
- Add Dockerfile.runtime (minimal debian + pre-built binary)
- Add Dockerfile.web-gateway-runtime (Node dashboard + pre-built binary)
- Keep Dockerfile.training via Kaniko (needs CUDA dev image for H100)
- Remove all SCCACHE_BUCKET build-args from service builds
- Use dir:// context for Kaniko (only sends build-out/ dir, not full repo)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Mass cleanup of crates/ml/:
- Delete 121 dead/broken/superseded example files (-44,098 lines)
- Delete 6 QAT test files (-4,201 lines) — QAT is disabled at runtime
(trainer.rs falls back to FP32 with warning)
- Delete 2 stale markdown files in examples/
- Delete orphaned src/bin/train_tft.rs (unimplemented stub)
- Clean up Cargo.toml: remove stale [[example]] entries, add missing ones
- Fix stale binary references in log_size_test.rs
- Add infra consistency test (35 checks across Dockerfile/train.sh/Cargo.toml)
Remaining examples (7): train_baseline_rl, train_baseline_supervised,
evaluate_baseline, hyperopt_baseline_rl, hyperopt_baseline_supervised,
download_baseline, cuda_test
All 2390 lib tests pass. All examples compile. 35/35 infra checks pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move 17 library crates into crates/, CLI binary into bin/fxt,
consolidate 10 test crates into testing/, split config crate
from deployment config files.
Root directory reduced from 38+ to ~17 directories.
All Cargo.toml paths and build.rs proto refs updated.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
