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>
Bring Branching Dueling Q-Network (Tavakoli 2018) to full Rainbow parity
with the existing GPU hotpath. 3 independent advantage heads (exposure=5,
order=3, urgency=3) decompose the 45-action space into learnable branches.
H1 - CUDA fallback: gate GpuExperienceCollector when use_branching=true
(fused kernel hardcodes NUM_ACTIONS=5, incompatible with 45 factored)
H2 - Per-branch C51 distributional: each branch outputs [batch, n_d, atoms]
log-softmax, loss = avg of D cross-entropies vs projected Bellman target
M1 - NoisyNet: MaybeNoisyLinear enum in branch heads, reset_noise/disable_noise
wired through select_action, compute_loss, and set_eval_mode
M2 - Regime-conditional IS weights: Trending=1.2, Ranging=0.8, Volatile=0.6
applied to branching loss via ADX/CUSUM features at state[40:41]
M3 - State dim alignment: align_dim_for_tensor_cores() in from_dqn_params()
for H100 HMMA dispatch (8-byte alignment)
L1 - Fill simulator: splitmix64 replaces golden ratio hash (chi-squared tested)
L2 - Hyperopt 29D: branch_hidden_dim [64,256] added to PSO search space
Config plumbing: branch_hidden_dim, v_min/v_max/num_atoms, use_distributional,
use_noisy, noisy_sigma_init all flow from DQNConfig → BranchingConfig.
10 files, +3207/-125 lines, 33 branching tests + 387 ml-dqn + 284 ml-core pass.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
training_dtype() now returns BF16 on all CUDA devices, enabling full
tensor-core utilization. F32 is used only at boundaries (scalar
extraction, loss computation, softmax). VRAM estimator updated to
account for BF16 byte sizes, C51/QR atoms, dueling streams, NoisyNet
param doubling, and GPU PER buffer pre-allocation.
Changes:
- mixed_precision.rs: training_dtype() returns BF16 on CUDA
- curiosity.rs: F32 cast before scalar extraction
- network.rs: F32 output at NetworkLayers forward boundary
- traits.rs: estimate_trial_vram_mb_full() with BF16-aware sizing
- dqn.rs adapter: uses full estimator with worst-case architecture
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>
Four monitoring functions (estimate_avg_q_value_with_early_stopping,
collect_qvalue_statistics, compute_q_gap_for_epoch, compute_per_action_q_values)
iterated over batch_sample.experiences to build CPU tensors for forward passes.
When GPU PER (GpuPrioritized) is active, experiences is always vec![] — all data
lives on GPU tensors in gpu_batch. This created zero-element tensors with non-zero
shapes, triggering CUBLAS_STATUS_INVALID_VALUE on the next forward pass.
Fix: all four functions now check for gpu_batch.states and use it directly,
falling back to CPU experiences only for non-GPU buffers. Also removes debug
eprintln probes, wires new_on_device for DQN/RegimeConditionalDQN construction,
and adds GPU-aware smoke tests (33 pass, 874 total, 0 failures).
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>
Move optimizers, gradient_accumulation, gradient_utils, cuda_compat,
tensor_ops, and gpu (device config, capabilities, memory profiling) to
ml-core. These are shared compute primitives used by all models.
Also commit module files for core types (common, config, error, model,
traits, types) that were moved from ml to ml-core in task 5a but left
staged without being committed.
Notable changes:
- resolve_batch_size() stays in ml (new batch_size_resolver module)
because it depends on memory_optimization::auto_batch_size which
has not yet moved to ml-core
- FactoredAction legacy bridge converted from inherent impl to
extension trait (FactoredActionLegacy) since FactoredAction is now
defined in ml-core, not ml
- candle-optimisers added to ml-core dependencies (needed by Adam)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move all inline type definitions from ml/src/lib.rs to ml-core:
MLError, MLResult, Trade, MarketRegime, HealthStatus, Features,
MLModel trait, ModelRegistry, ParallelExecutor, LatencyOptimizer,
TrainingMetrics, ValidationMetrics, InferenceResult, ModelMetadata.
Dedup: consolidate ConfigError{reason}/ConfigurationError(msg) into
single ConfigError(String) tuple variant (was 2 variants, 174 refs).
Cleanup: convert create_hft_* free functions to associated methods
(HFTPerformanceProfile::ultra_low_latency(), ParallelExecutor::hft()).
ml facade re-exports via `pub use ml_core::*`.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
PSO can now tune sharpe_weight in [0.0, 0.5] instead of hardcoded 0.3.
This lets hyperopt discover the optimal Sharpe ratio blending weight
in the composite reward signal per-symbol.
Also updates docstring to reflect current C1-C4 search space state.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The DQN hyperopt adapter has an explicit trades_data_dir field from CLI
args, but load_ofi_features() was only using the derived sibling path.
Now uses the explicit field when available, falling back to derivation.
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>
Port factorized Gaussian noise exploration and 51-atom categorical value
distribution to the CUDA experience collection kernel, fixing a correctness
bug where GPU Q-values were wrong when use_distributional=true (production
default) due to misinterpreted distributional weight shapes.
- D5: NoisyNet factorized noise (Box-Muller + f(x)=sign(x)*sqrt(|x|)) on
all 6 dueling layers, online network only — target stays deterministic
- D6: C51 distributional dueling forward with per-action atom softmax,
RMSNorm after shared/value/advantage layers, correct [51,128]/[255,128]
weight interpretation
- RmsNormWeightSet extraction and post-epoch sync (GPU-to-GPU)
- Fix get_effective_epsilon() to report actual 2% noisy floor instead of 0.0
- Proportional diversity entropy penalty (continuous gradient vs cliff)
- 2758 tests passing, 0 failures
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The walk-forward backtest in the DQN hyperopt adapter constructed batch
tensors with raw state_dim (51/43) but the model expects aligned
state_dim (56/48). This caused shape mismatch errors during backtest
evaluation: matmul [1024, 51] vs [56, 1024].
Fix: use align_dim_for_tensor_cores() and zero-pad each state vector
before tensor construction, matching the same pattern used in
compute_loss_internal.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move BF16 tensor core alignment from per-forward-pass allocation to
data pipeline boundaries. On H100, state_dim 43→48 and 51→56 (8-aligned)
so cuBLAS dispatches HMMA instructions instead of falling back to scalar FMA.
Architecture:
- Trainer computes aligned state_dim at source (align_dim_for_tensor_cores)
- GPU path: DqnGpuData.pad_state_tensor() pads once at upload boundary
- CPU path: train_batch() fold zero-pads Experience.state vectors
- Networks receive pre-aligned tensors — zero per-step overhead
All state_dim defaults updated to aligned values (43→48, 51→56).
Removed pad_to_aligned() from all network forward() methods.
2758 tests pass, 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both adapters now call load_ofi_features_parallel from mbp10_loader
instead of duplicating the rayon + streaming logic inline.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three fixes:
- Streaming MBP-10 parser (parse_mbp10_streaming): computes OFI inline
during decode — eliminates Vec<Mbp10Snapshot> allocation (41.9M clones)
- Parallel file processing: rayon par_iter across 9 MBP-10 files
(778s sequential → ~90s expected on H100 24-core)
- Fix hardcoded state_dim=54 in walk-forward backtest tensor creation
that caused panic "range end index 55296 out of range for slice of
length 52224" — now uses dynamic state_dim (43 or 51 with OFI)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: preload_data() called loader.ofi_features.take() on the internal
DQN trainer, but load_training_data() only loads OHLCV bars — it never
populates ofi_features. The OFI loading is done by the hyperopt adapter's
own load_ofi_features() method.
Fixes:
- preload_data() now calls self.load_ofi_features() directly
- load_ofi_features() uses self.mbp10_data_dir when set (was hardcoded ../mbp10)
- input_dim pre-computation uses OFI-aware size (51 when enabled, 43 otherwise)
- CI 'latest' package: delete-then-upload to avoid GitLab duplicate file issue
- Warn when mbp10_data_dir is set but no OFI features loaded
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The hyperopt preload_data() created a loader with default hyperparams
(mbp10_data_dir=None), so MBP-10/trades data was never loaded. Each
trial then set mbp10_data_dir → state_dim=51, but the preloaded data
had no OFI features → shape mismatch [128,43] vs [51,1024].
- Pass mbp10_data_dir/trades_data_dir to preload hyperparams
- Extract ofi_features from loader after preload
- Store as preloaded_ofi_features: Option<Arc<Vec<[f64;8]>>>
- Inject into each trial's DQNTrainer before training
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add --mbp10-data-dir and --trades-data-dir CLI args to
hyperopt_baseline_rl binary so hyperopt trials can use real
order book and trade data for VPIN/Kyle's Lambda features.
- DQNTrainer: add mbp10_data_dir/trades_data_dir fields + with_ofi_data_dirs() builder
- DQNHyperparameters: pipe through from trainer instead of hardcoded None
- download-trades-job: fix nodeSelector to ci-compile-cpu (platform pool full)
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>
1. Walk-forward windows: replaced 3 non-overlapping with sliding (50% overlap, ~5 windows).
Aggregation changed from mean-0.5*std to median-0.5*IQR for outlier robustness.
2. Composite score: tanh normalization prevents Calmar ratio scale dominance
(0.02% drawdowns → values in thousands drowning out Sharpe/Sortino).
3. Q-value overestimation: new Prometheus gauge foxhunt_training_q_overestimation_ratio,
warning log when ratio>10 or q_mean>5, adaptive tau doubles when Q-mean growth>0.5/epoch
(capped at 0.01), decays back when stable.
2742 tests pass, 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- C1: Narrow V_min/V_max from [-15,+15] to [-2,+2] for C51 atom resolution
- C2: Remove exploration stacking (28D→26D), keep NoisyNet only
- C3: Fix entropy regularizer for 5-action space, remove 2x/3x multipliers
- H1: Narrow gamma to [0.88, 0.96] (avoid overnight gap discounting)
- H2: Raise tau lower bound to 0.005 (target net tracks in short trials)
- H3: Cap kelly_fractional at 0.75 (no full-Kelly ruin)
- M1: Enable QR-DQN when num_atoms > 100 (hyperopt-toggled)
- M3: Add Q-value gap logging (Q_best - Q_second_best) per epoch
2735 tests pass, 0 clippy warnings
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The objective function was using buy/sell/hold percentages from training
metrics instead of the backtest. This caused the optimizer to receive stale
action distribution signals that diverged from actual backtest behavior.
Added buy_action_pct/sell_action_pct/hold_action_pct to BacktestMetrics,
counted during the multi-window backtest loop, and used in extract_objective
when backtest metrics are available.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Patience-based early stopping (WAVE 24) was not gated by early_stopping_enabled,
causing ALL hyperopt trials to terminate early and return penalty metrics with
backtest_metrics: None — no backtest ever ran, producing FALLBACK OBJECTIVE 44.6.
Also removes eval_softmax_temp from search space (29D→28D) since backtest uses
greedy argmax, widens gamma range (0.95-0.99→0.90-0.999) for better TPE signal,
and updates all tests to match.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Greedy argmax legitimately converges to 1-2 actions when the model
is confident. The hard gate (unique_actions < 3 → 8.0 penalty) was
blocking ALL trials from using real backtest metrics, forcing FALLBACK.
Changes:
- Remove hard diversity short-circuit gate (unique_actions < 3)
- Reduce soft diversity penalty from 3.0 to 0.8 max (mild TPE signal)
- Fix early_stopping_enabled: false (was self.epochs > 12)
- Update test to validate mono-action policies produce good objectives
Multi-window backtest (3 windows, mean - 0.5*std) already handles
phantom Sharpe from single-bucket flukes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
apply_position_mask() in factored_q_network.rs looped 0..45 against a
5-wide Q-values tensor — would panic at runtime. Changed to 0..5 using
ExposureLevel::from_index() directly. Updated stale "45 actions" comments
in 5 files (dqn.rs, reward.rs, hyperopt/adapters/dqn.rs, curriculum.rs).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
A: Switch backtest eval from Gumbel softmax to greedy argmax (batch_greedy_actions)
so hyperopt Sharpe reflects the agent's actual learned policy, not noisy sampling.
C: Disable reward normalization (enable_normalization=false). EMA normalizer with
±3.0 clipping was flattening the reward landscape, preventing the agent from
distinguishing large winners from scratch trades.
D: Wire tx_cost_bps (0.1 bps for IBKR ES) through to EvaluationEngine via
new_with_fee_rate(). Previously hardcoded at 15 bps (150x mismatch with actual
commission costs), massively penalizing every trade in backtest.
E: Scale PnL reward by agent's target exposure in calculate_pnl_reward().
Previously, a Short100 action received POSITIVE reward when market went up
(pct_return ignored position direction). Now: reward = pct_return × exposure.
2735 tests pass, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Early stopping with 8-epoch trials returns penalty metrics (no backtest),
causing ALL trials to hit FALLBACK OBJECTIVE = 44.6 regardless of actual
model quality. The Sharpe was 1.36-1.61 but natural fluctuation triggered
"Sharpe worsening" at epoch 5, killing the backtest evaluation.
Fix: disable early stopping when epochs ≤ 12. With ~90s per trial, running
all 8 epochs is cheap. Long training runs (50 epochs) still use it.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The C2 fix from the previous session was too aggressive — it eliminated
ALL exploration mechanisms simultaneously:
1. epsilon forced to 0.0 when noisy nets active (select_action)
2. count bonus removed from Q-value computation (metrics only)
3. noisy_epsilon_floor config field declared but never read
This left noisy nets as the sole exploration mechanism, which produces
perturbations too small to overcome Q-value gaps (A4=0.12 vs others≈0.02).
Result: 20/20 hyperopt trials hit fallback objective with 1/5 diversity.
Fixes:
- select_action: use noisy_epsilon_floor (not 0.0) as effective_epsilon
when noisy nets active — guarantees minimum random action rate
- select_action: re-enable UCB count bonus on Q-values before argmax
(both IQN and standard paths) for directed exploration
- select_action_with_confidence: same fixes for consistency
- trainer: set epsilon to noisy_epsilon_floor (not 0.0) at init
- hyperopt: widen noisy_epsilon_floor range from [0.0, 0.05] to
[0.03, 0.15] with default 0.05
Exploration now has two complementary mechanisms:
- noisy_epsilon_floor: random actions feed diverse replay buffer
- count bonus: UCB term biases greedy selection toward under-visited actions
- noisy nets: weight perturbation adds stochasticity to Q-values
select_action_inference (production) is unchanged — pure exploitation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Critical bug: all 3 DQN action selection methods (select_action,
select_action_with_confidence, select_action_inference) used
FactoredAction::from_index() which maps indices 0-4 to exposure_idx=0
(Short100) via division by 9. This is the root cause of action
diversity collapse during both training and production inference.
Fix: ExposureLevel::from_index() + OrderRouter::route_default() in all
DQN paths. Also fixes hyperopt objective thresholds (<10 → <3 for
5-action degenerate detection), stale defaults/comments, integration
test configs.
Files: dqn.rs (3 methods), trainer.rs (validation + select_action),
hyperopt/adapters/dqn.rs (thresholds), dqn_model.rs (comments),
train_baseline_rl.rs (default), reward.rs (comment),
dqn_integration.rs + ensemble_integration.rs (num_actions).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- agent.rs: select_action_factored() now uses ExposureLevel::from_index()
+ OrderRouter::route_default() instead of FactoredAction::from_index().
Previously, indices 0-4 mapped to all-Short100 variants in the 45-action
space — now correctly maps to 5 distinct exposure levels.
- hyperopt: plateau_window .max(3) → .max(5) to prevent premature early
stopping with short trial epochs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: 45 factored actions (5 exposure × 3 order × 3 urgency) caused
reward degeneracy — 9 actions per exposure level produced nearly identical
rewards since order type/urgency had 1000-4000x weaker signal than PnL.
This collapsed action diversity as DQN couldn't differentiate actions.
Changes:
- DQN now outputs 5 Q-values (Short100, Short50, Flat, Long50, Long100)
- New OrderRouter deterministically maps exposure → (order_type, urgency)
based on spread and volatility microstructure signals
- PPO retains full 45-action space (separate CUDA constants DQN_NUM_ACTIONS
vs PPO_NUM_ACTIONS)
- CUDA kernels: DQN diversity entropy uses 5 categories, PPO keeps 45
- Phase B: pnl_history cleared per epoch so Sharpe reflects current epoch
(was accumulating across all epochs, causing frozen Sharpe metric)
24 files, 2728 tests pass, 0 clippy warnings
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sharpe-based early stopping kills every hyperopt trial at epoch 4
because compute_epoch_financials() is deterministic (greedy argmax on
fixed validation data) — the model doesn't change enough in 8 short
epochs to shift any argmax decisions, making Sharpe bit-identical
across epochs and triggering plateau detection immediately.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Multi-window backtest: splits validation data into 3 non-overlapping
windows and aggregates with mean(Sharpe) - 0.5*std(Sharpe), penalizing
inconsistency and reducing overfit to a single data segment.
Top-K ensemble: hyperopt now emits top_k_params (top 5 trials) in JSON
output. train_baseline_rl gains --ensemble-top-k flag to train multiple
models per fold from different hyperopt configs, saving checkpoints as
dqn_ensemble_{k}_fold_{n}.safetensors.
Workflow template: adds ensemble-top-k parameter (default 5) and passes
--ensemble-top-k to the train-best step.
2720 tests pass, 0 clippy warnings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
