jgrusewski 796fbd01cf fix: set c51_alpha_max=0.5 — prevent C51 gradient convergence on H100 ...

H100 fold 2 showed grad_norm→0 when C51 fully replaced MSE (alpha=1.0).
MSE provides a non-vanishing gradient floor. Default 0.5 = 50/50 blend.
Hyperopt can search [0.3, 0.9].

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

2026-04-09 20:02:48 +02:00

..

asset_selection

…

backtesting

…

benchmark

refactor: remove entire CPU training path — 5,307 lines of dead code

2026-04-02 09:06:23 +02:00

checkpoint

…

cuda_common

…

cuda_pipeline

fix(critical): IQN d_h_s2 is f32 not bf16 — remove bogus bf16_to_f32 cast

2026-04-09 19:59:26 +02:00

data_loaders

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

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

data_pipeline

…

data_validation

…

deployment

…

diffusion

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

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

dqn

refactor: remove entire CPU training path — 5,307 lines of dead code

2026-04-02 09:06:23 +02:00

ensemble

fix(4branch): num_actions 9→3, IQN total_branch_actions includes branch 3

2026-04-08 14:34:04 +02:00

evaluation

…

explainability

…

features

test: feature range validation — all 42 features bounded + no dead constants

2026-04-03 18:30:44 +02:00

flash_attention

…

gpu_benchmarks

…

hyperopt

feat: add c51_alpha_max to cap MSE→C51 blend and prevent gradient starvation

2026-04-09 15:42:41 +02:00

integration

…

kan

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

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

labeling

…

liquid

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

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

mamba

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

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

microstructure

…

model_registry

…

observability

…

paper_trading

…

ppo

…

regime

…

regime_detection

…

registry

…

risk

…

security

…

stress_testing

…

tests

…

tft

fix: resolve all CI test failures — TFT, DQN pipeline, benchmarks

2026-03-21 09:09:51 +01:00

tgnn

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

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

tlob

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

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

trainers

fix: set c51_alpha_max=0.5 — prevent C51 gradient convergence on H100

2026-04-09 20:02:48 +02:00

training

…

transformers

…

universe

…

validation

refactor: remove entire CPU training path — 5,307 lines of dead code

2026-04-02 09:06:23 +02:00

xlstm

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

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

.gitignore

…

batch_processing.rs

…

benchmarks.rs

…

bridge.rs

…

data_loader.rs

…

feature_cache.rs

feat: cache key includes symbol + data_source (prevents cross-instrument collisions)

2026-04-04 09:41:53 +02:00

fxcache.rs

feat: shared discover_and_load() — strict key match, no fallback

2026-04-04 09:51:28 +02:00

inference_validator.rs

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

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

inference.rs

…

lib.rs

refactor: remove entire CPU training path — 5,307 lines of dead code

2026-04-02 09:06:23 +02:00

model_factory.rs

…

model_loader_integration.rs

…

model_registry.rs

…

portfolio_transformer.rs

…

preprocessing.rs

…

training_pipeline.rs

…

training_profile.rs

feat: add c51_alpha_max to cap MSE→C51 blend and prevent gradient starvation

2026-04-09 15:42:41 +02:00

training.rs

…

walk_forward.rs

perf: zero-copy fold loop — fxcache to GPU once, index slices per fold

2026-04-02 22:46:30 +02:00