jgrusewski 907b5f3355 fix: cross-stream race in PER update + pre-allocate scratch buffers ...

The replay buffer used its own CUDA stream for seg_tree_update, but
the td_errors buffer was computed on the trainer's stream. Without
synchronization, the replay stream could read incomplete data — causing
a hang on H100 where streams execute truly concurrently.

Fix: update_priorities_gpu now accepts an optional ext_stream parameter.
The fused training path passes the trainer's main stream, ensuring all
GPU work (td_errors computation + seg_tree_update) runs on the same
stream with implicit ordering.

Also pre-allocate update_td_f32, update_batch_max, update_max_merge
scratch buffers at construction — eliminates 3 cuMemAlloc calls per
training step that could trigger implicit device synchronization.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-06 12:36:15 +02:00

..

common

…

edge_cases

…

fixtures

…

integration

…

regression

…

unit

…

validation

…

ab_testing_integration.rs

…

action_loader_real_csv_test.rs

…

action_masking_smoke_test.rs

…

activation_tests.rs

fix(tests): update integration tests for f32 pipeline + adam_epsilon

2026-03-29 21:32:04 +02:00

adx_features_test.rs

…

barrier_optimization_test.rs

…

bayesian_changepoint_test.rs

…

bessel_correction_test.rs

…

bug16_portfolio_features_test.rs

…

bug18_cash_reserve_solvency_test.rs

…

bug20_portfolio_normalization_test.rs

…

bug21_bug22_bug23_compilation_fixes_test.rs

…

bug24_bug25_compilation_fixes_test.rs

…

bug28_unused_import_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

cash_accounting_fix_test.rs

…

cash_reserve_requirement_test.rs

…

checkpoint_test.rs

…

configurable_capital_test.rs

…

contract_multiplier_test.rs

…

cusum_test.proptest-regressions

…

debug_position_delta_test.rs

…

diffusion_integration.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

dqn_accumulation_convergence_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

dqn_action_collapse_fix_test.rs

fix: CQL alpha test assertion matches default (1.0)

2026-04-04 01:04:45 +02:00

dqn_action_dependent_reward_test.rs

…

dqn_action_masking_integration_test.rs

…

dqn_action_position_sign_convention_test.rs

…

dqn_activity_penalty_fix_test.rs

…

dqn_adapter_paths_test.rs

…

dqn_advanced_metrics_integration_test.rs

…

dqn_barrier_integration_debug_test.rs

…

dqn_c51_shape_validation_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

dqn_diagnostic_logging_test.rs

…

dqn_diversity_penalty_test.rs

…

dqn_dueling_batched_wave62_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

dqn_early_stopping_termination_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

dqn_episode_boundaries_test.rs

…

dqn_epsilon_decay_validation_test.rs

…

dqn_feature_cache_test.rs

chore: rename test_data/mbp10 to mbp10-fixtures, remove uncompressed .dbn

2026-04-01 01:13:19 +02:00

dqn_feature_defaults_test.rs

refactor: remove 8 always-on use_ booleans — features are mandatory

2026-03-25 09:45:54 +01:00

dqn_gradient_accumulation_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

dqn_gradient_collapse_root_cause_test.rs

…

dqn_gradient_dtype_simple_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

dqn_hft_barriers_test.rs

…

dqn_hyperopt_json_export_test.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

dqn_hyperopt_test.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

dqn_hyperparams_kelly_fields_test.rs

…

dqn_inference_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

dqn_kelly_regime_integration_test.rs

…

dqn_logit_clipping_bug12_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

dqn_long_training_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

dqn_pnl_calculation_tests.rs

…

dqn_rainbow_config_test.rs

…

dqn_rainbow_test.rs

…

dqn_regime_conditional_integration_test.rs

…

dqn_reward_calculation_test.rs

…

dqn_tensor_shape_validation.rs

…

dqn_trainer_integration_tests.rs

refactor: remove 8 always-on use_ booleans — features are mandatory

2026-03-25 09:45:54 +01:00

dqn_trainer_p1_tests.rs

cleanup: remove ALL 32 enable_* feature flags — all features unconditional

2026-03-31 10:56:20 +02:00

dqn_training_pipeline_test.rs

fix: profile is single source of truth, q_clip wired as guard, fewer walk-forward folds

2026-04-04 15:25:02 +02:00

dqn_transaction_cost_bug2_test.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

dqn_transaction_costs_test.rs

…

dqn_use_double_dqn_test.rs

refactor: remove 8 always-on use_ booleans — features are mandatory

2026-03-25 09:45:54 +01:00

dropout_scheduler_tests.rs

fix(tests): update integration tests for f32 pipeline + adam_epsilon

2026-03-29 21:32:04 +02:00

ensemble_4_models_integration.rs

…

ensemble_hot_swap_test.rs

…

ensemble_hyperopt_tests.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

ensemble_inference_integration_test.rs

…

evaluation_net_pnl_bug3_test.rs

…

feature_extraction_46_test.rs

…

feature_normalization_test.rs

…

fxcache_roundtrip_test.rs

feat: has_ofi flag in fxcache header (explicit, no zero-detection)

2026-04-04 09:48:38 +02:00

gpu_backtest_validation.rs

fix(tests): update integration tests for f32 pipeline + adam_epsilon

2026-03-29 21:32:04 +02:00

gpu_kernel_parity_test.rs

fix(tests): update integration tests for f32 pipeline + adam_epsilon

2026-03-29 21:32:04 +02:00

gpu_per_integration_test.rs

fix: cross-stream race in PER update + pre-allocate scratch buffers

2026-04-06 12:36:15 +02:00

gradient_accumulation_tests.rs

refactor: remove 8 always-on use_ booleans — features are mandatory

2026-03-25 09:45:54 +01:00

huber_loss_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

hyperopt_action_masking_test.rs

cleanup: remove ALL 32 enable_* feature flags — all features unconditional

2026-03-31 10:56:20 +02:00

hyperopt_early_stopping_infrastructure_test.rs

…

hyperopt_kelly_params_test.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

hyperopt_paths_test.rs

…

hyperopt_price_extraction_test.rs

…

hyperopt_tft_early_stopping_test.rs

…

imbalance_bars_test.rs

…

inference_engine_test.rs

…

integration_ppo_ensemble.rs

…

kan_integration.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

kelly_criterion_integration_test.rs

…

liquid_cfc_training_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

liquid_networks_test.rs

…

liquid_nn_training_tests.rs

…

log_size_test.rs

…

lr_scheduler_tests.rs

…

mamba2_accuracy_fix_test.rs

…

mamba2_early_stopping_test.rs

fix: wire real GPU backprop + SPSA gradients, fix checkpoint loading, eliminate candle from examples

2026-03-20 15:41:42 +01:00

mamba2_gradient_extraction_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

mamba2_hyperopt_edge_cases.rs

…

mamba2_hyperopt_p0_p1_fixes.rs

…

mamba2_weight_update_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

mbp10_parsing_test.rs

chore: rename test_data/mbp10 to mbp10-fixtures, remove uncompressed .dbn

2026-04-01 01:13:19 +02:00

microstructure_features_test.rs

…

model_registry_checkpoint_test.rs

…

model_validation_comprehensive.rs

…

multi_cusum_test.rs

…

ofi_features_test.rs

fix: update 9 test files for 14D DQNParams restructure

2026-03-31 01:30:33 +02:00

ood_input_handling_tests.rs

…

pages_test_test.rs

…

paper_trading_integration_test.rs

…

partial_reversal_support_test.rs

…

partial_reversal_validation.rs

…

pnl_realism_tests.rs

…

portfolio_integration_tests.rs

…

portfolio_tracker_reset_test.rs

…

portfolio_value_normalization_test.rs

…

ppo_45_action_network_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_action_masking_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_checkpoint_roundtrip_test.rs

…

ppo_circuit_breaker_tests.rs

…

ppo_continuous_bounds_test.rs

…

ppo_continuous_transaction_costs_tests.rs

…

ppo_dual_phase_backtest_tests.rs

…

ppo_entropy_regularization_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_factored_action_tests.rs

…

ppo_hidden_state_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_huber_loss_validation.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_hyperopt_validation_test.rs

…

ppo_long_training_test.rs

…

ppo_lstm_architecture_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_lstm_training_loop_tests.rs

…

ppo_portfolio_tracker_tests.rs

…

ppo_position_limits_tests.rs

…

ppo_recurrent_integration_tests.rs

…

ppo_recurrent_performance_tests.rs

refactor: remove gpu_n_episodes override — auto-scale from VRAM everywhere

2026-04-02 14:21:45 +02:00

ppo_reward_normalizer_tests.rs

…

ppo_sequence_batching_multi_episode_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

ppo_step_counter_fix_test.rs

…

ppo_stress_testing_tests.rs

…

ppo_transaction_costs_tests.rs

…

ppo_validation_real_data_test.rs

…

preprocessing_bessel_integration.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

preprocessing_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

preprocessing_validation_tests.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

price_validity_tests.rs

…

production_trainer_adaptive_features_integration_test.rs

…

production_training_smoke_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

pso_budget_calculation_test.rs

…

q_value_constraint_test.rs

…

rainbow_loss_shape_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

real_data_helpers.rs

…

real_data_pipeline_test.rs

…

recovery_tests.rs

…

regime_transition_features_test.rs

…

risk_action_masking_test.rs

…

risk_position_limit_integration_test.rs

…

run_bars_test.rs

…

safety_comprehensive_test.rs

…

sample_weights_test.rs

…

scatter_add_gradient_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

smoke_test_real_data.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

softmax_exploration_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

streaming_pipeline_edge_cases.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

supervised_gpu_smoke_test.rs

fix: supervised GPU smoke test dimension mismatches

2026-03-21 09:14:53 +01:00

target_update_tests.rs

fix(tests): update integration tests for f32 pipeline + adam_epsilon

2026-03-29 21:32:04 +02:00

test_dbn_sequence_256_features.rs

…

test_dqn_cuda_device.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

test_extract_256_dim_features.rs

…

test_grn_weight_initialization.rs

…

test_scatter_source_gradients.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

test_sell_bug.rs

…

test_streaming_loader.rs

…

test_var_gradient_flow.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

test_var_source_gradients.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

tft_adapter_paths_test.rs

…

tft_causal_masking_validation.rs

…

tft_hyperopt_real_metrics_test.rs

…

tft_inference_latency_benchmark.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

tft_lru_cache_test.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00

tft_quantile_loss_validation.rs

…

tft_real_dbn_data_test.rs

…

tft_test.rs

…

training_chaos_tests.rs

…

training_edge_cases.rs

…

transaction_cost_application_test.rs

…

transition_matrix_test.rs

…

triple_barrier_test.rs

…

volatility_epsilon_test.rs

…

wave16_checkpoint_regression_test.rs

…

wave16q_ohlcv_mutation_test.rs

refactor: all train() and load_training_data() callers pass explicit symbol

2026-04-04 10:02:37 +02:00

wave16r_position_limits_test.rs

…

wave16s_price_validation_test.rs

…

wave_d_e2e_nq_fut_225_features_enhanced_test.rs

…

wave_d_e2e_nq_fut_225_features_test.rs

…

wave_d_edge_cases_test.rs

…

xlstm_integration.rs

fix: migrate 44 test files from Candle to native CUDA — zero test compile errors

2026-03-19 10:02:26 +01:00