GPU experience collector:
- Add GpuHardwareInfo with SM count detection from device name lookup
(H100=132, A100=108, L40S=142, RTX 4090=128, etc.)
- optimal_n_episodes() scales to GPU: sm_count × 2 warps × 32 threads,
capped by 15% free VRAM budget, 256-aligned for block scheduling
- Remove hardcoded .min(256) cap in trainer — auto-scales from 128 to 8192
- MAX_EPISODES_LIMIT raised from 256 to 4096
Numerical stability:
- BF16 cross-entropy epsilon: 1e-8 → 1e-4 (below BF16 precision floor
1e-8 rounds to zero, making log-stabilization a no-op → -inf → NaN)
- Add log_probs.clamp(-20, 0) guard against -inf × 0 = NaN in loss
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
BF16's 7-bit mantissa overflows on exp(50+) in softmax → Inf → NaN.
Cast distributional logits to F32 before softmax in both dueling and
rainbow network heads, then cast back to original dtype.
Increase default batch_size 128→1024 (standard), 128→256 (conservative),
512→2048 (aggressive) to saturate H100 tensor cores. AutoBatchSizer
already caps to VRAM ceiling for smaller GPUs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three root causes for "no metrics" on the training dashboard:
1. Dashboard template variables ($model, $fold) sourced from
foxhunt_training_current_epoch which isn't emitted until the first
epoch completes. Switch to foxhunt_training_step which fires from
step 500 onward.
2. train.sh pod template missing Prometheus annotations
(prometheus.io/scrape, port, path). Also add the
app.kubernetes.io/component label to the eval manifest so
evaluation pods are discoverable too.
3. DQN and PPO trainers only called set_epoch() at the END of each
epoch. Move the call to the TOP of the epoch loop so the gauge
exists from the first training iteration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three fixes validated by 20/20 hyperopt trials on H100 (zero OOM):
1. Workspace default-features: ml-core, ml-dqn, ml-ppo, ml-supervised
workspace deps now have default-features=false. Prevents cudarc
(which requires nvcc) from leaking into CPU service builds via
Cargo feature unification. CI compile-services was failing with
"Failed to execute nvcc: No such file or directory" (exit 101).
2. BF16 comparison fix: Candle's gt()/le() don't support BF16 operands.
Cast ADX/CUSUM features to F32 before threshold comparison in
regime classification. Previous approach (cast threshold to BF16)
failed due to Candle broadcast_as reverting dtype.
3. CI pipeline: expand ML change detection to all 14 sub-crates,
add component:compile labels for sccache network policy matching,
bump training runtime to CUDA 12.6 + Ubuntu 24.04 (glibc 2.39).
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>
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>
Replace expensive CPU↔GPU data transfers with GPU-native operations:
- check_gradients_finite: sum_all() scalar readback (4B per tensor)
replaces flatten_all()+to_vec1() that copied entire gradients to CPU
(up to 200MB per step across 9 call sites in DQN+PPO)
- Huber loss: affine() replaces 3× Tensor::from_vec(vec![const; N])
CPU Vec allocations in the inner loss computation loop
- update_priorities_gpu: per-index slice_scatter (~1KB DMA) replaces
full-buffer to_vec1()+from_vec() roundtrip (1.2MB DMA per step)
- PER sampling: single from_vec + GPU to_dtype replaces duplicate
from_vec calls and CPU type conversion roundtrips
- RegimeConditional: log warning on silent GPU batch drops
822 tests pass (350 ml-dqn, 274 ml-core, 198 ml-ppo), 0 clippy.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three fixes for GPU PER hot path:
1. IQN quantile loss used empty CPU weights Vec instead of GPU-resident
weights_tensor_cached — caused CUDA_ERROR_ILLEGAL_ADDRESS from
uninitialized GPU memory. Now uses cached GPU tensor matching C51
and standard DQN paths.
2. GpuPrioritized add()/add_batch() replaced with StagedGpuBuffer:
add() stages on CPU (Vec::push, zero GPU ops), sample() batch-flushes
staging→GPU in one DMA before sampling. Production path (insert_batch_tensors)
bypasses staging entirely — GPU→GPU with zero CPU.
3. All 9 ML sub-crates default to cuda feature so `cargo test -p ml-dqn`
exercises GPU code paths on CUDA workstations. CI service crates use
default-features=false, unaffected.
Test results: 350 passed (was 343+7 failed), 0 failed, 1 ignored.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Wire GpuReplayBuffer into DQN constructor with OOM fallback to CPU PER
- Fix P0: GPU PER priorities never updated in single-batch train_step()
(result.indices was empty CPU Vec; GPU tensors td_errors_gpu/indices_gpu
were silently dropped)
- Fix IS-weight ordering: apply weights AFTER Huber loss, not before
(weighting before nonlinear Huber shifts quadratic/linear regime boundary)
- Fix IQN CVaR: add .contiguous() before sort_last_dim (341/341 tests pass)
- Defer loss scalar readback to after backward pass (piggyback on grad flush)
- Add next_states to ExperienceBatch with episode-aware shift computation
- Add insert_batch_tensors() for direct GPU tensor insertion into replay buffer
- Make Q-value estimation periodic (every 50 steps) to reduce forward passes
- Fix CPU fallback path types (u8 action, i32 fixed-point reward, timestamp)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ml-supervised, ml-ensemble, ml-labeling, ml-explainability, ml-hyperopt
all had default = ["cuda"] which pulled cudarc into the CPU services
build, causing compile-services to fail with "nvcc not found".
Changed all to default = [] and wired ml/Cargo.toml cuda feature to
propagate to all 8 sub-crates (was only 3: ml-core, ml-dqn, ml-ppo).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove cuda from default features in ml-core, ml-dqn, ml-ppo, ml
- Propagate cuda feature from ml → ml-core/ml-dqn/ml-ppo
- CI compile-training already uses --features ml/cuda explicitly
- Fix MaxDD log format: {:.1}% → {:.3}% (was rounding 0.033% to 0.0%)
- Suppress unused_labels/unused_variables warnings for cfg(cuda) code
- Add CALLBACK_ENDPOINT env to ml-training-service deployment
- Fix Grafana active_workers query to use sum() with fallback
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
GPU kernel produced raw percentage PnL rewards (~3.5e-4) which were
~10,000x smaller than Q-values (~7.0), making rewards invisible in the
Bellman backup. Q-values froze at initialization.
Port the CPU RewardNormalizer algorithm (EMA mean/variance, z-score
normalization, [-3,3] clamp) directly into the CUDA kernel hot path
as per-thread register state. Each thread maintains its own running
mean/variance with configurable decay rate (reward_norm_alpha, default
0.01 = ~100-step window). EMA resets on episode boundaries.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Move UnifiedTrainable trait, TrainingMetrics, CheckpointMetadata, and
checkpoint helpers from ml to ml-core (zero new dependencies — ml-core
already had candle-core + serde_json)
- Wrap Mamba2SSM in Mamba2TrainableAdapter to satisfy orphan rule (trait
in ml-core, type in ml-supervised — all other 9 models already used
wrapper pattern)
- Make Mamba2SSM::validate() pub for cross-crate adapter access
- Delete 5 permanently disabled deployment modules (cfg(any()) — never
compiled): registry, hot_swap, validation, monitoring, endpoints
(-5,842 lines)
- Delete 2 undeclared dead files in training/: dqn_trainer.rs,
transformer_trainer.rs (-138 lines)
- Fix pre-existing compute_loss test shape mismatch in mamba adapter
UnifiedTrainable in ml-core unblocks future trainers/ extraction (17.8K
lines) since model-specific trainers can now depend on ml-core for the
trait without pulling in the full ml monolith.
14 files changed, +128 -6,497 (net -6,369 lines)
Tests: 274 ml-core + 948 ml = 1,222 passed, 0 failed
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 24 PPO source files + flow_policy/ from ml into standalone
ml-ppo crate. The ml crate's ppo module is now a thin re-export layer
(`pub use ml_ppo::*`) plus two bridge files (trainable_adapter.rs,
stress_testing.rs) that depend on ml-internal types.
Key changes:
- ml-ppo: 25 modules (incl flow_policy subdir), 198 tests, standalone
- ml: depends on ml-ppo, re-exports via ppo/mod.rs
- Import rewrites: crate::common::action → ml_core::action_space,
crate::dqn::{mixed_precision,xavier_init} → ml_core::*,
crate::gradient_accumulation → ml_core::gradient_accumulation
- ml tests: 1929 pass (down from 2127 — 198 moved to ml-ppo)
- Workspace: 0 errors
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move 53 DQN source files + gpu_replay_buffer from ml into standalone
ml-dqn crate. The ml crate's dqn module is now a thin re-export layer
(`pub use ml_dqn::*`) plus two bridge files (trainable_adapter.rs,
stress_testing.rs) that depend on ml-internal types.
Key changes:
- ml-dqn: 45 modules, 334 tests, compiles standalone
- ml: depends on ml-dqn, re-exports via dqn/mod.rs
- DQN.config: pub(crate) → pub for cross-crate access
- DQNAgent checkpoint methods moved into ml-dqn
- gpu_replay_buffer moved from cuda_pipeline/ to ml-dqn
- 8 dead-code files removed (never declared in mod.rs)
Net: -30,648 lines from ml crate. Workspace: 0 errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move Sharpe ratio metrics and SIMD performance module to ml-core.
These only depend on MLError which is already in ml-core.
Add approx = "0.5" to ml-core dev-dependencies for Sharpe tests.
Skip observability/ (needs prometheus dep — stays in ml).
Test counts: 271 ml-core + 2487 ml = 2758 total, 0 failures
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move 6 shared modules (~2.3K lines) from ml to ml-core:
- trading_action.rs (TradingAction enum)
- action_space.rs (action masking, re-exports)
- xavier_init.rs (Xavier/Glorot weight initialization)
- mixed_precision.rs (AMP utilities, FP16/BF16)
- order_router.rs (deterministic order routing)
- portfolio_tracker.rs (portfolio state tracking)
With TradingAction now in ml-core alongside FactoredAction, the
FactoredActionExt extension trait is eliminated entirely —
from_trading_action()/to_trading_action() become inherent methods
on FactoredAction. This removes the need for `use FactoredActionExt`
imports in reward.rs, gae.rs, and trajectories.rs.
Test counts: 250 ml-core + 2508 ml = 2758 total, 0 failures
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Move memory_optimization/ (4.6K lines, 7 files) from ml to ml-core
- Move batch_size_resolver.rs to ml-core (now both deps in same crate)
- Add tempfile to ml-core dev-dependencies (qat tests)
- Re-export both modules from ml facade
- Keep checkpoint/ in ml (model_implementations.rs has model-specific deps)
185 ml-core tests + 2573 ml tests = 2758 total, 0 failures
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Move safety/ module (6.5K lines) from ml to ml-core (self-contained)
- Keep security/ in ml (depends on ensemble::EnsembleDecision)
- Move From<ProductionTrainingError> for MLSafetyError to ml (cross-crate)
- Rename Real-prefixed inference types to proper names:
RealInferenceError → InferenceError
RealInferenceConfig → InferenceConfig
RealPredictionResult → InferencePrediction
RealNeuralNetwork → NeuralNetwork
RealMLInferenceEngine → MLInferenceEngine
- Re-export safety from ml facade for backward compatibility
129 ml-core tests + 2629 ml tests = 2758 total, 0 failures
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>
Move mixed_precision, xavier_init, portfolio_tracker, action_space,
order_router from dqn/ to ml/src/ root. Extract TradingAction enum
from dqn/agent.rs to standalone trading_action module. Re-export
from dqn/mod.rs for backward compatibility.
These modules are shared infrastructure used by PPO, trainers,
hyperopt, and ensemble — not DQN-specific.
Test results: 2757 passed, 0 failed, 25 ignored (unchanged).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three bugs in the GPU experience collection hot path:
1. gpu_batch_to_experiences() had hardcoded state_dim=43 but the CUDA kernel
outputs states at the ALIGNED dimension (56 with OFI, 48 without). After
sample 0, every replay buffer entry had corrupted state vectors — the
network was learning from garbage data.
2. GPU path never called monitor.track_reward(), so mean_reward was always
reported as 0.0 in epoch logs despite the agent generating real rewards.
3. Action tracking was double-counted (direct array write + track_action_by_exposure),
inflating diversity metrics by 2x. Consolidated into single bounded call.
Also adds missing app.kubernetes.io/component label to job-template.yaml
so Prometheus training-pods scrape job discovers training pods.
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 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>
from_dqn_params() was stripping the last hidden_dim, creating only 1
shared layer (shared_0) when hidden_dims=[256,256]. The CUDA experience
kernel hardcodes 2 shared layers and gpu_weights.rs expects shared_1.*
weight keys. This caused "Missing weight: shared_1.weight" every epoch,
forcing CPU fallback for all experience collection on H100.
Fix: use all hidden_dims as shared layers. Default [256,256] now creates
shared_0 + shared_1, matching the CUDA kernel architecture exactly.
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>
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>
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>
GPU experience collector was falling back to CPU because it required a
curiosity VarMap, but curiosity_weight=0.0 means the module is never
created. Fix: make curiosity optional (CuriosityWeightSet::zeros() for
GPU buffers, curiosity_scale=0.0 in kernel config). Also require
explicit binary-tag SHA in Argo training workflow (no "latest" fallback).
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>
Additional padding in get_q_values and convert_to_state — these are
currently unused in the training pipeline but would crash if called.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Validation, action selection, and Q-value monitoring paths were using
raw state dimensions (51) while the model expected aligned dims (56).
Training epoch 1 passed because the GPU pipeline pads correctly, but
validation crashed: shape mismatch [1000,51] vs [56,1024].
Fixed: validation batch, select_actions_batch CPU fallback,
estimate_avg_q_value, compute_q_gap — all now zero-pad to aligned dim.
Also fixed model size estimation log to show aligned dims.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add QuestDB ILP sink for training metrics, update Prometheus scrape
configs, and fix network policies for monitoring stack connectivity.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
