jgrusewski befa7e2f55 perf(gpu): shared memory weight caching, GPU reductions, fix staging flush dim ...

CUDA kernel: cooperative_load_tile() + matvec_leaky_relu_shmem() cache weight
matrices in shared memory (SRAM), reducing HBM traffic 32x (1.76TB→55GB).
Block size increased to 256 threads for better SHMEM utilization.

Trainer hot-path: replace to_vec2 full-tensor CPU readbacks with GPU-native
flatten_all/min/max/mean_all (4 scalar readbacks). Merge two Q-diagnostic
forward passes into one. Deferred max_priority flush (1 GPU→CPU sync/epoch
instead of ~8/batch).

Fix StagedGpuBuffer::flush() dimension mismatch: used raw experience
state.len() (45) instead of gpu.state_dim() (48 aligned), causing
slice_scatter crash [500,48] vs [n,45] when CPU experiences flush into
GPU-aligned replay buffer.

Fix hardcoded state_dim=48 in optimal_n_episodes() — now uses dynamic
state_dim from network config (correct for OFI-enabled 56-dim states).

Dynamic episode auto-scaling guard: only auto-scale when gpu_n_episodes≥128
(production), respecting explicit test overrides (gpu_n_episodes=2).

Tests: 874 ml + 392 ml-dqn = 1266 pass, 0 fail.

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

2026-03-09 22:24:33 +01:00

..

backtesting

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

common

perf(metrics): expose step-level DQN metrics to Prometheus gauges

2026-03-09 18:44:29 +01:00

config

refactor: rename JWT issuer foxhunt-api-gateway → foxhunt-api, drop compat aliases

2026-03-05 00:22:04 +01:00

ctrader-openapi

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

data

fix(ml): streaming parallel OFI + fix hardcoded state_dim=54 in backtest

2026-03-07 23:28:18 +01:00

database

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

market-data

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

ml

perf(gpu): shared memory weight caching, GPU reductions, fix staging flush dim

2026-03-09 22:24:33 +01:00

ml-asset-selection

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

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

ml-backtesting

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

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

ml-checkpoint

refactor(ml): remove dead code across 8 sub-crates — delete 742 lines

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

ml-core

perf(gpu): dynamic episode scaling + BF16 cross-entropy stabilization

2026-03-09 21:16:47 +01:00

ml-data

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

ml-data-validation

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

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

ml-dqn

perf(gpu): shared memory weight caching, GPU reductions, fix staging flush dim

2026-03-09 22:24:33 +01:00

ml-ensemble

fix(dqn): IQN GPU PER weights, staged GPU buffer, CUDA default in all ML crates

2026-03-08 22:31:49 +01:00

ml-explainability

fix(dqn): IQN GPU PER weights, staged GPU buffer, CUDA default in all ML crates

2026-03-08 22:31:49 +01:00

ml-features

fix: resolve rebase artifacts — import paths and MLError variant

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

ml-hyperopt

fix(dqn): eliminate OOM in hyperopt by gating GPU features on small GPUs

2026-03-09 16:35:38 +01:00

ml-labeling

fix(dqn): IQN GPU PER weights, staged GPU buffer, CUDA default in all ML crates

2026-03-08 22:31:49 +01:00

ml-observability

refactor(ml): remove dead code across 8 sub-crates — delete 742 lines

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

ml-paper-trading

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

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

ml-ppo

fix(dqn): cast F32 constants to input dtype for BF16 mixed precision on H100

2026-03-09 17:13:46 +01:00

ml-regime

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

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

ml-regime-detection

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

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

ml-risk

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

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

ml-security

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

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

ml-stress-testing

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

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

ml-supervised

fix(dqn): IQN GPU PER weights, staged GPU buffer, CUDA default in all ML crates

2026-03-08 22:31:49 +01:00

ml-universe

refactor(ml): remove dead code across 8 sub-crates — delete 742 lines

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

ml-validation

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

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

model_loader

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

risk

refactor: remove trading_engine dep from ml and risk crates

2026-03-05 23:19:38 +01:00

risk-data

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

storage

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

trading_engine

refactor: replace all api_gateway/web-gateway references across codebase

2026-03-05 00:18:58 +01:00

trading-data

docs: create/update README.md for all 17 crates

2026-03-01 22:47:39 +01:00

training_uploader

feat: implement all 15 fxt CLI commands with real gRPC calls

2026-03-03 22:16:35 +01:00