jgrusewski bf87a8d0bb perf: kan_grad_reduce — batch-parallel 2-phase (8ms×208 → <0.1ms×208) ...

Replace single-phase one-thread-per-param serial loop (batch_size iterations
per thread) with kan_grad_reduce_p1 (block sums, grid=(ceil(B/256),total_params),
shared mem) + kan_grad_reduce_p2 (warp-shuffle final reduce, grid=(total_params)).
Allocate partials scratch [ceil(B/256)*total_params] for trunk + 4 branches.
Update CublasBackwardSet constructor signature and both call sites.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-19 12:49:03 +02:00

..

asset_selection

…

backtesting

…

benchmark

refactor: collapse 9-level to 7-level ExposureLevel — eliminate degenerate Flat variants

2026-04-11 11:54:09 +02:00

checkpoint

cleanup: delete dead CPU replay buffers, DQNAgentType::Standard, add CUDA event profiling

2026-04-11 00:13:09 +02:00

cuda_common

…

cuda_pipeline

perf: kan_grad_reduce — batch-parallel 2-phase (8ms×208 → <0.1ms×208)

2026-04-19 12:49:03 +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

…

dqn

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

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

ensemble

WIP: training quality fixes + VarStore elimination (INCOMPLETE — needs xavier_init_params_buf)

2026-04-11 21:03:05 +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: cost-driven hold timing — replace min_hold_bars with learned cost signals

2026-04-18 01:04:11 +02:00

integration

cleanup: delete dead CPU replay buffers, DQNAgentType::Standard, add CUDA event profiling

2026-04-11 00:13:09 +02:00

kan

…

labeling

…

liquid

…

mamba

…

microstructure

…

model_registry

…

observability

…

paper_trading

…

ppo

refactor: collapse 9-level to 7-level ExposureLevel — eliminate degenerate Flat variants

2026-04-11 11:54:09 +02:00

regime

…

regime_detection

…

registry

…

risk

…

security

…

stress_testing

…

tests

…

tft

…

tgnn

…

tlob

…

trainers

fix: epoch metrics from fused_ctx accumulators (every step, not guard-only)

2026-04-19 10:23:51 +02:00

training

…

transformers

…

universe

…

validation

refactor: collapse 9-level to 7-level ExposureLevel — eliminate degenerate Flat variants

2026-04-11 11:54:09 +02:00

xlstm

…

.gitignore

…

batch_processing.rs

…

benchmarks.rs

cleanup: delete dead CPU replay buffers, DQNAgentType::Standard, add CUDA event profiling

2026-04-11 00:13:09 +02:00

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

fix: fxcache version display + export FXCACHE_VERSION + cleanup dead code

2026-04-17 02:12:33 +02:00

inference_validator.rs

…

inference.rs

…

lib.rs

cleanup: remove half crate dependency from entire workspace

2026-04-10 18:54:14 +02:00

model_factory.rs

…

model_loader_integration.rs

…

model_registry.rs

…

portfolio_transformer.rs

…

preprocessing.rs

…

training_pipeline.rs

…

training_profile.rs

feat: cost-driven hold timing — replace min_hold_bars with learned cost signals

2026-04-18 01:04:11 +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