jgrusewski e89fce5eb2 fix(cuda): per-thread EMA reward normalization in GPU experience kernel ...

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>

2026-03-08 16:49:11 +01:00

..

asset_selection

refactor(ml): extract universe, backtesting, asset-selection into sub-crates

2026-03-08 15:17:22 +01:00

backtesting

refactor(ml): extract universe, backtesting, asset-selection into sub-crates

2026-03-08 15:17:22 +01:00

benchmark

fix(dqn,ppo): prevent 45-action collapse with entropy, exploration, and monitoring fixes

2026-03-03 02:04:22 +01:00

checkpoint

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

cuda_common

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

cuda_pipeline

fix(cuda): per-thread EMA reward normalization in GPU experience kernel

2026-03-08 16:49:11 +01:00

data_loaders

feat(ml): real trade data pipeline for VPIN/Kyle's Lambda + offline RL

2026-03-07 19:14:46 +01:00

data_pipeline

refactor(ml): move core types to ml-core + dedup MLError (5a)

2026-03-08 15:13:41 +01:00

data_validation

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

deployment

refactor(ml): move UnifiedTrainable to ml-core + delete 6K dead deployment code

2026-03-08 15:17:22 +01:00

diffusion

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

dqn

refactor(ml): extract DQN module into ml-dqn crate (task 6)

2026-03-08 15:16:08 +01:00

ensemble

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

evaluation

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

explainability

refactor(ml): extract regime-detection, explainability, paper-trading

2026-03-08 15:17:22 +01:00

features

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

flash_attention

refactor(ml): move safety/ to ml-core + rename Real-prefixed types (5c)

2026-03-08 15:13:41 +01:00

gpu_benchmarks

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

hyperopt

fix: resolve rebase artifacts — import paths and MLError variant

2026-03-08 15:21:46 +01:00

integration

refactor(ml): move core types to ml-core + dedup MLError (5a)

2026-03-08 15:13:41 +01:00

kan

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

labeling

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

liquid

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

mamba

refactor(ml): move UnifiedTrainable to ml-core + delete 6K dead deployment code

2026-03-08 15:17:22 +01:00

microstructure

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

model_registry

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

observability

refactor(ml): extract observability, stress-testing, security into sub-crates

2026-03-08 15:17:22 +01:00

paper_trading

refactor(ml): extract regime-detection, explainability, paper-trading

2026-03-08 15:17:22 +01:00

ppo

refactor(ml): extract PPO module into ml-ppo crate (task 7)

2026-03-08 15:16:08 +01:00

regime

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

regime_detection

refactor(ml): extract regime-detection, explainability, paper-trading

2026-03-08 15:17:22 +01:00

registry

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

risk

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

security

refactor(ml): extract observability, stress-testing, security into sub-crates

2026-03-08 15:17:22 +01:00

stress_testing

refactor(ml): extract observability, stress-testing, security into sub-crates

2026-03-08 15:17:22 +01:00

tests

refactor(ml): remove all legacy naming and deprecated wrappers

2026-03-08 15:13:41 +01:00

tft

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

tgnn

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

tlob

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

trainers

refactor(ml): move compute primitives to ml-core (5b)

2026-03-08 15:13:41 +01:00

training

refactor(ml): move UnifiedTrainable to ml-core + delete 6K dead deployment code

2026-03-08 15:17:22 +01:00

transformers

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

universe

refactor(ml): extract universe, backtesting, asset-selection into sub-crates

2026-03-08 15:17:22 +01:00

validation

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

xlstm

refactor(ml): extract 8 supervised models into ml-supervised crate (task 8)

2026-03-08 15:16:08 +01:00

.gitignore

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

batch_processing.rs

refactor(ml): move core types to ml-core + dedup MLError (5a)

2026-03-08 15:13:41 +01:00

benchmarks.rs

refactor(ml): consolidate duplicate Rainbow DQN agent implementations

2026-03-04 18:27:42 +01:00

bridge.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

data_loader.rs

fix(ml): BUG #40 — epsilon stuck at 1.0 with noisy nets (100% random actions)

2026-03-04 23:13:25 +01:00

examples.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

feature_cache.rs

fix(ml): DQN hyperopt overhaul — DSR reward, dead features, C51/noisy/network fixes

2026-03-07 12:08:39 +01:00

inference_validator.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

inference.rs

refactor(ml): move safety/ to ml-core + rename Real-prefixed types (5c)

2026-03-08 15:13:41 +01:00

lib.rs

refactor(ml): split monolith into 9 sub-crates + delete dead code

2026-03-08 15:17:22 +01:00

model_factory.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

model_loader_integration.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

model_registry.rs

refactor: restructure repo — crates/, bin/, testing/ layout

2026-02-25 11:56:00 +01:00

portfolio_transformer.rs

feat(ml): BF16 VarBuilder, checkpoints, and training tensors for PPO

2026-03-03 16:16:07 +01:00

preprocessing.rs

feat(ml): expand DQN hyperopt search space 27D→31D + GPU-batched eval inference

2026-03-05 20:45:13 +01:00

training_pipeline.rs

feat(ml): comprehensive GPU saturation audit — 58 fixes across all 10 models

2026-03-02 15:41:23 +01:00

training.rs

refactor: delete 4,800 lines of dead ML code and remove MLFeatureExtractor dependency

2026-03-01 18:29:43 +01:00

walk_forward.rs

feat(ml): real trade data pipeline for VPIN/Kyle's Lambda + offline RL

2026-03-07 19:14:46 +01:00